{"paper_id":"346d9ef2-b62f-4f30-9c64-2802a4bcf9b7","body_text":"Deep-learning based quantitative evaluation of postoperative atelectasis following right upper lobectomy | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Deep-learning based quantitative evaluation of postoperative atelectasis following right upper lobectomy Devanish N. Kamtam, Guiseppe M. Facchi, Nicole Lin, Lillian L. Tsai, and 8 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7768040/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 30 Apr, 2026 Read the published version in npj Digital Medicine → Version 1 posted 15 You are reading this latest preprint version Abstract Objective: Chronic postoperative atelectasis of lung lobes is an occasional complication following thoracic surgery, particularly middle lobe atelectasis following right upper lobectomy (RULobectomy). Existing methods of grading atelectasis are typically manual, subjective, and not scalable. We aimed to develop and validate an automated, deep learning–based volumetric framework to quantify and grade postoperative atelectasis using pre- and post-operative CT scans. Methods: We retrospectively included all patients who underwent RULobectomy in our institution from 2008 to 2023 who had available pre-operative and 6-month postoperative CT scans. We trained two separate nnU-Net v2 segmentation models for preoperative and postoperative lobar anatomy followed by volumetric quantification of the right middle lobe (RML), right lower lobe (RLL), and total lung volume. Atelectasis severity in the RML was independently graded by two surgeons using a standardized, 5-point radiological scale (none, minimal, subsegmental, segmental, lobar). The association between volume metrics and clinical atelectasis severity was evaluated using both the original 5-point scale and a pooled 3-point scale (none, minimal–subsegmental, segmental–lobar). Results: 236 patients comprised the study cohort. The pre- and postoperative models achieved high segmentation accuracy (mean Dice scores: 0.98 ± 0.02 and 0.99 ± 0.00, respectively). Median (IQR) RML volume loss progressively increased with higher atelectasis grades, from − 4.6 mL (-78.5, 59.0) in grade 0 to − 317.8 mL (-440.7, -194.8) in grade 4 atelectasis (p < 0.001). Normalized RML/right lung (RL) and RML/total lung (TL) volume ratios showed statistically significant differences across the pooled atelectasis grades (p < 0.001). Conversely, normalized RLL volumes increased with worsening RML atelectasis (p < 0.001), suggesting compensatory hyperinflation. Conclusions: We demonstrate the feasibility and clinical relevance of deep learning–based volumetric assessment of atelectasis after RULobectomy. This automated pipeline with its open-source model enables reproducible quantification of atelectasis through lobar volume loss and may serve as a scalable tool for clinical and research applications involving atelectasis. Health sciences/Diseases Health sciences/Medical research artificial intelligence deep learning lobectomy middle lobe atelectasis segmentation volumetry nnU-Net Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction Atelectasis – partial or complete collapse of a lung lobe resulting from alveolar deflation or inadequate expansion – is one of the most common pulmonary complications following surgery 1 . It is particularly common following cardiothoracic surgery, with rates of 30–72% 2,3 , versus 3.2% following non-cardiothoracic surgery 4 . Atelectasis can be classified into obstructive (resorptive), compressive, and adhesive types, based on underlying mechanisms such as internal airway obstruction, external compression, or surfactant dysfunction. Collapse of lung lobes reduces the alveolar surface area available for gas exchange, leading to impaired oxygenation, hypoxemia, and, in severe cases, respiratory failure. Further, completely atelectatic lung lobes and segments have a propensity to become infected. While early postoperative atelectasis is often self-limited, complete lobar collapse and ongoing, substantial sublobar collapse can lead to significant morbidity, especially following lung resections where the remaining lung must compensate for the loss of resected parenchyma. Postoperative atelectasis is typically identified through radiographic imaging, with chest X-rays for preliminary assessment, followed by CT scans for diagnostic confirmation and more detailed assessment of the severity of atelectasis. To facilitate consistent evaluation of atelectasis —albeit largely in non-postoperative scenarios—several studies have proposed standardized grading criteria 5 , some of which have also been validated against clinical outcomes 6 – 8 . Early objective approaches to grading atelectasis primarily relied on Hounsfield Unit (HU) measurements 9 – 12 . CT slice intensity values between − 500 and + 100 HU (vs. < -500 HU as aerated lung) have been used to identify poorly aerated regions reflecting atelectasis 9 – 12 , and these metrics demonstrated good correlation with clinical outcomes. While effective within the lung parenchyma, this approach becomes less reliable when atelectasis extends to the lung periphery, as the collapsed lung border may be difficult to distinguish from adjacent mediastinal structures on CT. And in complete lobar collapse, the proportion of collapsed voxels may be outweighed by the volume of the remaining aerated lung, leading to underestimation of atelectasis severity. Moreover, despite HU standardization (–1000 HU - air, 0 HU - water), the HU thresholds may differ by scanner type/ imaging acquisition protocols (Supplementary Fig. 1). More recently, other methods, such as the BEST-CT 13 , have been used to quantify atelectasis as a percentage of total lung volume. This method divides axial CT slices into grids and classifies each segment into one of ten status categories, including atelectasis, the percentage of which is then calculated across the entire scan. While this methodology has been validated in two clinical contexts—cystic fibrosis 14 and bronchiectasis 13 — the grading remains entirely manual and time-intensive. Recently, less time-intensive scoring systems like the ASSESS criteria have been introduced in post-interventional settings such as following bronchoscopy under general anesthesia, grading atelectasis by its extent on CT from the posterior chest wall to the anterior vertebral border. Their applicability, however, is limited to dependent dorso-caudal lung regions. Overall, existing methods of grading atelectasis remain constrained by their manual, condition-specific, subjective nature, with significant interrater variability. This limits their generalizability, scalability, and reproducibility. Deep learning may enable more robust and automated grading of atelectasis that is reproducible across diverse clinical contexts, imaging protocols, and healthcare systems—provided such variability is adequately represented in the training data. Atelectasis can be graded either by segmenting and quantifying the reduction in volume of the collapsed regions of the lung 5 , 9 or, alternatively, by segmenting and quantifying volume of the aerated regions that represent the remaining functional lung. Prior studies 15 , 16 have confirmed that quantitative volumetric analysis of lung aeration—whether by CT or MRI 16 , 17 —provides objective metrics that can serve as reliable proxies for assessing the degree of atelectasis. Hence, we aimed to use automated volumetric analysis of aerated lung regions to quantify and grade the severity of atelectasis. Given its higher incidence following thoracic surgery, we evaluated atelectasis following pulmonary lobectomy, using right upper lobectomy (RULobecttomy)—the most frequently performed lobectomy—as a representative model. This was also an ideal model because chronic right middle lobe (RML) atelectasis, often referred to as right middle lobe syndrome, is a recognized long-term complication after right upper lobectomy 18 . Given this known risk, we specifically targeted the quantification of middle lobe volume loss and its correlation with physician-assigned atelectasis grades. We leveraged paired preoperative and six-month-postoperative CT scans from patients who had undergone RULobectomy to develop and validate an automated, deep learning–based volumetric pipeline to quantify/grade atelectasis. Our aim was to establish a reproducible and clinically interpretable framework for quantifying lobar volume changes and grading postoperative atelectasis and to evaluate its clinical relevance by comparing volumetry-derived atelectasis metrics with physician-assigned grades in a large postoperative cohort. This approach may offer a scalable tool for both retrospective research and prospective clinical risk stratification, with potential applicability to other lobectomies and clinical contexts in which atelectasis occurs. Methods Study design This study was a single-center retrospective analysis. The Checklist for Artificial Intelligence in Medical Imaging (CLAIM-2024) was used in study design and implementation 19 . The CLAIM checklist is modeled after the Standards for Reporting of Diagnostic Accuracy Studies (STARD) guidelines and is part of the EQUATOR network, ensuring best practices in reporting and facilitating the translation of AI into clinical practice. The study was approved by the Stanford University School of Medicine Institutional Review Board (IRB), which waived the requirement for informed consent as the study involved only retrospective chart review (IRB-70048). All procedures were carried out in accordance with the Declaration of Helsinki ethical standards. Patient cohort and clinical data We retrospectively reviewed all consecutive patients who underwent a right upper lobectomy in the Division of Thoracic Surgery at Stanford University Hospital from January 2008 through December 2023. Patients were identified using Stanford’s customized version of the Society of Thoracic Surgery (STS) General Thoracic Database and the STAnford Research Repository (STARR) clinical data warehouse. Patients without a preoperative CT scan within 6 months prior to surgery or a postoperative CT scan performed at approximately 6 months (range: 3–9 months) after surgery were excluded. Patients with scans of axial slice thickness > 3 mm were also excluded. Baseline demographic and clinical data were extracted from the databases and the electronic medical record. Follow-up data were collected for up to 6 months post-operatively. Outcomes Atelectasis of the RML was graded using a 5-point scale: 0 = none, 1 = minimal/linear, 2 = subsegmental, 3 = segmental, and 4 = near-total or total lobar collapse. CT scans were reviewed using a lung window setting (window width: 1500 HU; window level: − 600 HU). Image evaluation was performed using 3D Slicer (version 8.5.1; MIT, Massachusetts, USA) 20 . Two board-certified surgeons independently reviewed and graded all scans, resolving discrepancies through consensus until full agreement was achieved. Imaging acquisition and radiological data Preoperative and postoperative diagnostic chest CT scans were acquired using multidetector CT scanners from multiple vendors, most commonly Siemens (Siemens Healthineers, Erlangen, Germany) and GE Medical Systems (GE Healthcare, Chicago, IL, USA), occasionally with Toshiba (Toshiba Medical Systems Corporation, Otawara, Japan) and Philips (Philips Healthcare, Best, The Netherlands). Scans were performed during end-inspiratory breath-hold when tolerated. A wide variety of reconstruction kernels was used across the cohort, with the most frequent being STANDARD, T20f, Tr20f, and T20s. Axial images were reconstructed with slice thicknesses ranging from 0.3 mm to 3 mm. CT acquisition protocols varied, with tube voltages ranging from 80 to 130 kVp (median [IQR]: preoperative – 120 [120–120]; postoperative – 120 [80–120]) and tube currents ranging from 10 to 660 mA (median [IQR]: preoperative – 30 [20–35]; postoperative – 35 [20–35]). Most preoperative and postoperative scans were performed without intravenous contrast. Training and inference data The dataset was divided into training and test sets ensuring that no patient had scans in both sets. A separate validation set was not required, as the self-configuring nnU-Net framework does not rely on manual hyperparameter tuning. The training pipelines for both the preoperative and postoperative segmentation models are illustrated in Fig. 1 . Conversion of DICOM files to NIfTI format de-identified the images by removing all protected health information from the image metadata. To ensure generalizability, the preoperative lobe segmentation model was trained using two datasets: an internal institutional dataset and an external public dataset (OttawaChestCT dataset) 21 . The internal Stanford dataset included a random subset of 41 CT scans (34 for training and 7 for test) selected from a total of 236 clinical scans. The OttawaChestCT dataset contributed 81 scans (59 for training and 22 for test) out of the original 100; 19 scans were excluded due to incomplete lobe annotations (i.e., missing one or more of the five lobe masks). In total, the combined dataset comprised 122 annotated CT scans, with voxel-wise labels for six classes: background, right upper lobe (RUL), RML, right lower lobe (RLL), left upper lobe (LUL), and left lower lobe (LLL). The training pipeline for the postoperative lobe segmentation model utilized only the internal institutional dataset, consisting of randomly selected 82 CT scans (67 for training and 15 for test) from the full cohort of 236 clinical scans. Each scan was manually annotated with four segmentation classes: background, RML, RLL, and left lung. Reference standard for segmentation masks Reference standard segmentation masks were generated using 3D Slicer (version 8.5.1; MIT, Massachusetts, USA) by D.N.K. for preoperative and postoperative CT scans in the internal Stanford dataset, while the original annotations provided by the dataset creators were used for the OttawaChestCT public dataset 21 . Model training This study utilized the nnU-Net v2 model, a self-configuring deep learning framework for biomedical image segmentation. Its predecessor, nnU-Net v1, was built upon the U-Net architecture and designed to automatically adapt its network architecture, preprocessing steps, and training pipeline to the specific characteristics of a given dataset—eliminating the need for manual tuning of architectural or hyperparameter settings. The nnU-Net v2 retains the original self-adapting architecture and automated training configuration of the nnU-Net framework, with enhancements focused on improved training efficiency and broader device compatibility. Detailed implementation of the model is presented in the original paper 22 , 23 . Model training was performed using a single NVIDIA H100-SXM5 GPU. The 3D full-resolution nnU-Net v2 model was trained separately for preoperative and postoperative CT datasets, with each training run spanning 5,000 epochs (Fig. 2 ). Training time per epoch ranged from 22 to 30 seconds, resulting in a total training duration of approximately 42 hours for each model. Input CT images were resampled to a median voxel spacing of 1.25×0.70×0.70 mm and normalized using CT-specific intensity scaling. Due to the large volumetric size of CT scans in the dataset (median shape 258×512×512), training was performed with a patch size of 96×160×160 voxels and a batch size of 2. A Leaky ReLU activation function was used in all the hidden layers. The model was trained using the Adam optimizer with an initial learning rate of 0.001, following a polynomial (poly) learning rate decay schedule throughout training. Combined Dice and cross-entropy loss was used to guide optimization. Deep supervision was enabled. The checkpoint with the highest pseudo-Dice coefficient—i.e., the approximate Dice score computed on training patches rather than on the full CT volume—across all lung lobes on the training dataset was selected for further evaluation on both the test set and the full clinical dataset. Due to the absence of reference standard masks for the full clinical dataset, predicted lobe segmentations were visually inspected to confirm adequate anatomical accuracy. Preprocessing, data augmentation, and postprocessing These were handled using the default nnU-Net pipeline. Preprocessing included resampling CT scans to the median dataset voxel spacing, instance normalization using foreground intensity statistics, and cropping to nonzero regions to remove background. Data augmentation was performed on-the-fly and included random spatial transformation (rotation, scaling, elastic deformation) and intensity augmentations (gamma, noise, blur). Postprocessing involved removing all but the largest connected component for each predicted lobe label, applied selectively based on whether it improved the Dice score compared to baseline segmentation. In addition to the default nnU-Net postprocessing, customized automated postprocessing was applied to further refine the predicted masks. This included removal of anatomically discordant contralateral lobe predictions and isolated 3D islands disconnected from the largest lobar component. These customized postprocessing steps were not part of the training pipeline but were implemented post-inference and resulted in a marginal improvement in overall Dice scores. Volumetric quantification of atelectasis Preoperative and postoperative lung volumes for the full clinical dataset were quantified using automatically segmented masks derived from model inference on the respective CT scans. The preoperative CT segmented regions included RUL, RML, RLL, LUL, and LLL while the postoperative regions included RML, RLL, and left lung. Voxel volumes were calculated using the dimensions of each voxel. Lobar and total lung volumes were computed by multiplying the voxel volume (converted from mm³ to mL) with the number of voxels in each label. A subset of the calculated volumes was cross validated against measurements obtained using the standard volume computation tool in 3D Slicer (version 8.5.1; MIT, Massachusetts, USA) to ensure accuracy. RML and RLL volumes were also normalized with right lung and total lung volumes to minimize interscan variability. Statistical analysis Descriptive statistics were used to summarize baseline demographic and clinical data. The normality of continuous data was evaluated using the Shapiro-Wilk test. Normally distributed continuous variables were reported as mean ± standard deviation (SD) and compared using unpaired t-tests, while non-normally distributed variables were reported as median (interquartile range, IQR) and compared using the Mann-Whitney U test. Categorical variables were presented as frequencies and percentages and compared using Pearson’s chi-square test (χ²) or Fisher’s exact test, as appropriate. The Jonckheere–Terpstra test was used to assess for ordered trends in all volume metrics across atelectasis severity grades. A two-tailed α < 0.05 was considered statistically significant. Bonferroni correction was applied for multiple testing. All analyses were conducted using SPSS software (version 29.0, IBM Corp., Armonk, NY, USA), R (version 4.4.2, R Foundation for Statistical Computing, Vienna, Austria), and Python (version 3.13, Python Software Foundation, Wilmington, DE, USA). Results Baseline Cohort Characteristics Out of the 438 patients who underwent right upper lobectomy, 236 patients were included in the study after excluding those without available preoperative or postoperative scans within the defined time windows. Baseline demographic and clinical characteristics are summarized in Table 1 . The median age was 69.2 years (IQR: 61.0–74.6) and 62.3% of patients were female. Common comorbidities included diabetes mellitus (17.8%), chronic obstructive pulmonary disease (16.5%), cerebrovascular disease (11.4%), and chronic kidney disease (11.0%). Table 1 Baseline demographic and clinical characteristics of all patients who underwent right upper lobectomy and those who were included in the final study cohort. Characteristics Right upper lobectomy cohort (n = 438) Study cohort (n = 236) Age (years) (Median, IQR) 69.7(62.7–75.0) 69.2(61.0-74.6) Sex Male 179(40.9%) 89(37.9%) Female 259(59.1%) 147(62.3%) Race White 282(64.4%) 144(61.0%) Black 7(1.6%) 5(2.1%) Asian 98(22.4%) 64(27.1%) Other 43(9.8%) 23(9.7%) Unknown 8(1.8%) 0(0%) Approach Open 117(26.7%) 50(21.2%) VATS 248(56.6%) 134(56.8%) Robot-assisted 73(16.7%) 52(22.0%) FEV 1 (%) (Median, IQR) 95(79.2–108) 97(80–109) Diabetes mellitus 64(14.6%) 42(17.8%) Chronic obstructive pulmonary disease 70(16.0%) 39(16.5%) Congestive heart failure 25(5.7%) 15(6.4%) History of cerebrovascular accident 45(10.3%) 27(11.4%) Chronic kidney disease 40(9.1%) 26(11.0%) Charlson-comorbidity score (Median, IQR) 6(4–7) 6(5–8) Operative Diagnosis Adenocarcinoma 329(75.1%) 191(80.9%) SCC 43(9.8%) 14(5.9%) NEC 16(3.7%) 6(2.5%) Mixed 10(2.3%) 3(1.3%) Metastases 17(3.9%) 11(4.7%) Granulomatous inflammation 7(1.6%) 3(1.3%) Other benign lesions 16(3.7%) 8(3.4%) Median time from preop scan to surgery (days) (Median, IQR) 44.5(24.0–74.0) Median time from surgery to postop scan (days) (Median, IQR) 196(173.2-207.7) Atelectasis on CT scan at 6 months 0 260(78.3%) 186 (78.8%) 1 48(14.5%) 32 (13.6%) 2 13(3.9%) 10 (4.2%) 3 6(1.8%) 6 (2.5%) 4 5(1.5%) 2 (0.8%) Radiological data characteristics The median time between preoperative CT scans and the surgery was 44.5 (24, 74) days and the median time between the surgery and postoperative CT scans was 196 (173.2, 207.7) days. The median slice thickness was 1.25 (1.0, 1.25) mm for preoperative and postoperative scans. RML atelectasis at 6 months was absent in 186 (78.8%) of patients, while grade 1, 2, 3, and 4 RML atelectasis were observed in 32 (13.6%), 10 (4.2%), 6 (2.5%), and 2 (0.8%) patients. Preoperative model performance On the training dataset (n = 93), Dice scores ranged from 0.97 ± 0.01 for the RML to 0.99 ± 0.00 for the RLL, with an overall mean Dice of 0.98 ± 0.01. The internal test dataset (n = 29) showed similarly strong performance, achieving an overall mean Dice score of 0.98 ± 0.02 after postprocessing (Table 2 ). The RML consistently yielded the lowest Dice scores across training and test datasets, likely reflecting the higher likelihood of incomplete right horizontal fissure vs. other fissures and smaller volume of this lobe. Table 2 Preoperative and postoperative lung segmentation model performance on CT scans, reported as Dice scores for each lung lobe and overall mean Dice score. Preoperative Lung lobes Training dataset (93) Internal test dataset (29) External validation dataset (55) Model inference After post-processing Model inference After post-processing Right upper lobe 0.98 ± 0.01 0.97 ± 0.03 0.97 ± 0.03 0.94 ± 0.07 0.95 ± 0.06 Right middle lobe 0.97 ± 0.01 0.96 ± 0.03 0.96 ± 0.03 0.86 ± 0.22 0.86 ± 0.22 Right lower lobe 0.99 ± 0.01 0.98 ± 0.04 0.98 ± 0.03 0.93 ± 0.12 0.94 ± 0.12 Left upper lobe 0.99 ± 0.00 0.98 ± 0.02 0.98 ± 0.02 0.93 ± 0.13 0.95 ± .13 Left lower lobe 0.99 ± 0.01 0.98 ± 0.03 0.98 ± 0.01 0.89 ± 0.18 0.91 ± 0.18 Overall Mean 0.98 ± 0.01 0.97 ± 0.02 0.98 ± 0.02 0.91 ± 0.16 0.92 ± 0.15 Postoperative Lung lobes Training dataset (67) Internal test dataset (15) External validation dataset Model inference After post-processing - - Right middle lobe 0.96 ± 0.1 0.98 ± 0.01 0.98 ± 0.01 - - Right lower lobe 0.98 ± 0.01 0.99 ± 0.00 0.99 ± 0.00 - - Left lung 0.99 ± 0.02 0.99 ± 0.00 0.99 ± 0.00 - - Overall Mean 0.98 ± 0.04 0.99 ± 0.00 0.99 ± 0.00 - - On the external test dataset of LOLA11 challenge (n = 55), baseline model performance showed moderate variability across lobes, with Dice scores ranging from 0.86 ± 0.22 for the RML to 0.94 ± 0.07 for the RUL. Postprocessing yielded minor improvements in the mean Dice score from 0.91 ± 0.16 to 0.923 ± 0.15. This score currently is the joint state-of-the-art in the most comprehensive and longest running lung lobe segmentation challenge (0.928 ± 0.15) 24 . Postoperative model performance On the training dataset (n = 67), Dice scores ranged from 0.96 ± 0.1 for the RML to 0.99 ± 0.02 for the left lung, with an overall mean Dice score of 0.98 ± 0.04 (Table 2 ). Internal testing (n = 15) yielded excellent performance, with an overall Dice score of 0.99 ± 0.00 after postprocessing. Among the individual lobes, the RML showed slightly lower Dice scores. Automated inference, volumetric analysis, and atelectasis grading in clinical CT dataset We applied our preoperative and post-operative model inference to the entire clinical CT dataset (n = 236). Eight postoperative scans (3.4%) were excluded from analysis owing to pleural pathologies, including effusion and pneumothorax. Owing to the lack of a reference standard in the postoperative dataset, all output masks were visually inspected, and manual corrections were required in 16 scans (6.8%) to ensure accurate lobar volumetry for atelectasis assessment. The median Dice score between the edited and original masks (n = 16) was 0.85 (IQR, 0.74–0.94). We then extracted lobar volumes and assessed longitudinal lobar volume change using various metrics (Table 3 ). Median (IQR) ΔRML volume change demonstrated a progressive decline with increasing RML atelectasis grades from − 4.6 mL (-78.5, 59.0) (no atelectasis) to − 317.8 mL (-440.7, -194.8) (grade 4 atelectasis) (p < 0.001). This pattern persisted across normalized metrics, with ΔRML/RL and ΔRML/TL volume changes showing increasingly negative values with increasing grades of atelectasis (p < 0.001). Importantly, RLL volume changes showed an opposite trend, highlighting the expected compensatory hyperinflation of the RLL with increasing degrees of RML atelectasis. ΔRLL/RL volumes increased with worsening RML atelectasis (e.g., median ΔRLL/RL − 31.2% to 50.5%, p < 0.001) (Table 3 ). However, the ΔRLL/TL did not show a similar trend (p = 0.79). While the overall trend of RML volume loss and compensatory ΔRLL/RL volume increase was significant, pairwise comparisons reached statistical significance only between grades 0–1 and occasionally in grades 1–2, but not between higher grades (2–3 or 3–4), likely due to low sample sizes in those categories (Fig. 3 A-E). Table 3 Quantitative assessment of lobar volume changes between preoperative and 6-month postoperative CT scans, stratified by radiographic atelectasis grade of 0–4. Volume changes are presented as absolute (mL) and percentage changes (%) relative to right lung (RL) and total lung (TL) volumes (Median, IQR). Metrics None (0) Minimal (1) Sub-segmental (2) Segmental (3) Lobar (4) p-value (for trend) Median RML volume change (mL) -4.6 (-78.5, 59.0) -74.3 (-146.5, -40.5) -189.9 (-296.2, -96.9) -244.9 (-289.6, -199.2) -317.8 (-440.7, -194.8) < 0.001 Median RML/RL volume change (%) 3.7 (1.1, 6.5) 0.7 (-2.3, 3.2) -4.5 (-8.6, -1.1) -9.2 (-12.1, -5.1) -9.5 (-12.1, -7.0) < 0.001 Median RML/TL volume change (%) 0.6 (-0.9, 1.8) -1.7 (-2.7, -0.2) -3.1 (-5.2, -1.5) -5.2 (-6.7, -3.5) -5.1 (-6.4, -3.7) < 0.001 Median RLL/RL volume change (%) 31.2 (27.5, 36.3) 34.0 (29.4, 37.1) 38.2 (36.0, 42.5) 42.7 (35.6, 47.1) 50.5 (49.2, 51.9) < 0.001 Median RLL/TL volume change (%) 11.6 (9.5, 14.0) 11.1 (5.8, 13.4) 12.4 (9.1, 15.3) 13.0 (6.0, 19.5) 17.5 (17.1, 18.0) 0.79 RML, right middle lobe; RLL, right lower lobe; RL, right lung; TL, total lung. While individual comparisons between consecutive atelectasis grades did not reach statistical significance, comparisons after pooling the grades of atelectasis as minimal-subsegmental (grades 1–2) and segmental-lobar collapse (grades 3–4) showed a statistically significant difference (Table 4 ). Notably, changes in ΔRML, ΔRML/RL, and ΔRML/TL between grades 0 vs 1–2 and 1–2 vs 3–4 were highly significant (p < 0.001). And while ΔRLL/RL across atelectasis grades also reached significance, ΔRLL/TL did not (Fig. 3 F-J). Table 4 Quantitative assessment of lobar volume changes between preoperative and 6-month postoperative CT scans, stratified by pooled atelectasis grades (grade 0, 1–2, and 3–4) Volume changes are presented as absolute (mL) and percentage changes (%) relative to right lung (RL) and total lung (TL) volumes (p-values for pairwise comparisons in Fig. 3 F-J) (Median, IQR). Metrics None (0) Minimal (1) AND Sub-segmental (2) Segmental (3) AND Lobar (4) p-value (for trend) Median RML volume change (mL) -4.4 (-78.6, 59.0) -93.7 (-168.6, -43.7) -244.9 (-318.0, -190.9) < 0.001 Median RML/RL volume change (%) 3.7 (1.1, 6.5) -0.5 (-3.5, 2.6) -9.2 (-12.9, -4.5) < 0.001 Median RML/TL volume change (%) 0.6 (-0.9, 1.8) -1.7 (-3.1, -0.4) -5.2 (-7.0, -3.3) < 0.001 Median RLL/RL volume change (%) 31.2 (27.5, 36.3) 34.3 (29.6, 38.5) 45.4 (40.2, 48.8) < 0.001 Median RLL/TL volume change (%) 11.6 (9.5, 14.0) 11.1 (6.1, 13.6) 17.2 (7.5, 18.9) 0.72 RML, right middle lobe; RLL, right lower lobe; RL, right lung; TL, total lung. Sensitivity analysis of segmentation accuracy and volumetric assessments after excluding RML wedge resections and replacing the original unedited outputs Excluding cases with additional RML wedge resections beyond the index RULobectomy procedure, which could confound the interpretation of lobar volume loss, resulted in similar directional trends in volume metrics, reinforcing the association between automated volume metrics and physician-assigned atelectasis severity grades (Supplementary Fig. 2). Moreover, analysis using the unedited segmentation outputs for the postoperative dataset yielded similar statistical differences in most volume metrics across both atelectasis and pooled atelectasis grades (Supplementary Fig. 3). Qualitative analysis Various grades of RML atelectasis are illustrated in Figs. 4 and 5 . Qualitative error analysis on segmentation errors was performed on CT scans where the Dice score was < 0.97. The most frequent errors were observed along the fissure between the RUL-RML and at the hilar junction of RML–RLL. Additional inaccuracies included over-segmentation of solid regions within dense consolidation areas, and under-segmentation of bullae (Fig. 6 ). A detailed qualitative error analysis was performed on the external validation dataset, which is the most diverse publicly available dataset of chest CT scans for lung lobe segmentation (Fig. 9). Representative errors over complex anatomical variations and challenging pathologies, such as scoliosis, chest wall deformities, severe COPD, large pleural effusions with near-complete lung collapse, and other abnormalities are illustrated in Fig. 7 . Discussion In this study, we developed and validated an automated, deep learning-based volumetric pipeline to quantify and grade atelectasis–one of the most common clinical problems in pulmonary medicine and surgery. We selected postoperative atelectasis of the middle lobe following right upper lobectomy as a representative model to demonstrate that atelectasis can be graded using lobar volumetry. Importantly, such an automated system to reliably quantify atelectasis would also likely be applicable to the broader assessment of atelectasis across diverse clinical scenarios. Although deep learning–based models have demonstrated high accuracy in delineating lobar anatomy with Dice scores of 92–96% 25–28 (Table 5 ), they have been trained on preoperative anatomically normal lungs, with limited validation in postoperative settings where anatomical distortion is common. To date, only one study assessed lobar segmentation before and after various lobectomies, but without examining associations with clinical outcomes 29 . Our models achieved high accuracy in delineating lobar anatomy, even in the anatomically distorted postoperative setting, and further demonstrated that volumetric reduction of the RML with compensatory hyperinflation of RLL correlated with increasing clinically determined grades of atelectasis. These findings remained consistent in sensitivity analyses. Together, the results support feasibility of automated lobar volumetry as an objective surrogate for grading atelectasis severity, with the caveat that some cases were excluded and some required manual adjustment. Table 5 Prior state-of-the-art lung lobe segmentation performance reported in the literature and on the publicly available LOLA11 challenge dataset comprising 55 chest CT scans. Author (Year) [citation] Right upper lobe Right middle lobe Right lower lobe Left upper lobe Left lower lobe Overall Doel et al. (2012) – Traditional CV methods 34 0.86 ± 0.15 0.55 ± 0.40 0.77 ± 0.34 0.88 ± 0.21 0.86 ± 0.24 0.79 ± 0.31 3D U-Net (2016–2020) 35 – 37 0.93 ± 0.07 0.84 ± 0.12 0.93 ± 0.04 0.92 ± 0.04 0.93 ± 0.04 0.91 ± 0.08 Ferreira et al. FRV-Net (2018) 38 0.94 ± 0.07 0.87 ± 0.12 0.94 ± 0.05 0.95 ± 0.03 0.95 ± 0.04 0.92 ± 0.02 Gerard et al. Series of 3D CNN (2019) 39 0.99 0.98 0.99 0.99 0.99 0.99 Xie et al. RU-Net (2020) 27 – (IoU) 0.95 ± 0.03 0.96 ± 0.03 0.96 ± 0.01 0.96 ± 0.02 0.96 ± 0.02 0.95 ± 0.03 Zheng et al. (2021) - Dual attention network 40 0.93 ± 0.01 0.90 ± 0.01 0.96 ± 0.01 0.94 ± 0.01 0.94 ± 0.01 0.93 ± 0.02 Zhang et al. (2021) 41 DenseVNet 0.96 0.92 0.96 0.94 0.93 0.94 Peng et al. (2022) 28 - Multi-feature fusion and Ensemble. 0.97 0.88 0.96 0.98 0.97 0.95 Bao et al. (2023) 42 Edge enhancement cascaded network 0.98 0.96 0.98 0.98 0.97 0.97 Nomura et al (2025) 43 TriSwinUNETR 0.93 0.85 0.95 0.97 0.96 0.93 LOLA11 lung lobe segmentation challenge 24 Van Rikxkoort et al. (2010) 44 Traditional CV 0.85 Lassen et al. (2013) 45 Traditional CV 0.88 Bragman et al. (2017) 25 Probabilistic model 0.91 ± 0.20 0.88 ± 0.24 0.93 ± 0.07 0.80 ± 0.23 0.91 ± 0.19 0.88 Imran et al. (2018) 26 PDV-Net 0.93 ± 0.07 0.86 ± 0.12 0.95 ± 0.03 0.93 ± 0.03 0.94 ± 0.03 0.92 ± 0.07 Current SOTA – nnU-Net (2025) 0.95 ± 0.08 0.86 ± 0.22 0.96 ± 0.06 0.95 ± 0.13 0.92 ± 0.19 0.928 ± 0.15 Ours (2025) 0.95 ± 0.06 0.86 ± 0.22 0.94 ± 0.13 0.95 ± 0.13 0.91 ± 0.17 0.923 ± 0.16 Despite claims of superior performance of novel architectures—transformer-based and Mamba-based–over CNNs for 3D medical segmentation, the CNN-based nnU-Net demonstrated the best Dice scores in a recent comprehensive benchmarking of six 3D medical datasets 23 . In addition to its robust and generalizable performance with relatively small datasets and its self-configuring architecture, the added advantages of lower VRAM consumption and shorter training times vs. other architectures influenced our choice of nnU-Net for model training. Table 5 lists other preoperative lung lobe segmentation models reporting Dice scores up to 0.99 but which lack validation on external datasets like LOLA11 24 . Moreover, our study demonstrates that deep learning models can robustly segment even complex, anatomically distorted postoperative lung anatomy following lobectomy. We have also demonstrated the ability of the models to generalize across vendors/scanners, acquisition protocols, and reconstruction kernels. While the postoperative model was generally robust, ~ 6% of scans required manual editing to ensure accurate volumetry—typically in cases with more severe anatomical distortion. This highlights the potential for improved performance through training on larger, diverse datasets. To our knowledge, this is the first study to apply DL–based analysis for grading postoperative atelectasis. While prior work has attempted binary classification of atelectasis presence 30 – 32 , such approaches are overly simplistic and fail to reflect the true spectrum of disease severity. Moreover, manual grading systems are prone to mild-moderate inter- and intra-rater variability, limiting their reliability and scalability. Our approach addresses these limitations by not only detecting presence of atelectasis as a “present/absent” decision, but also assessing its severity in a more objective, continuous, and reproducible manner by quantifying lobar volume loss. While our study focused only on RML atelectasis, this approach is readily applicable to other lobes and clinical scenarios where serial imaging is available. It is important to consider a few clinical implications of this volumetric approach. While subtle streak-like atelectasis (grade 1) may not lead to overt volume loss, we observed significant differences in volume metrics between grade 0–1, suggesting that early changes can also be quantitatively captured. On a different note, our model segments lung lobes with all included structures such as tumor masses or areas of consolidation. And notably, unlike atelectasis, tumor masses or consolidation from pneumonia typically do not lead to a substantial reduction in lobar volume. Therefore, the volumetric approach may offer the potential to differentiate atelectasis from other causes of increased opacity, such as pneumonia. This distinction, however, would require more detailed validation. This volumetric approach for grading atelectasis also aligns with the principles of clinically explainable AI (xAI) as it yields outputs that are directly interpretable by clinicians, radiologists, or even technicians. This is unlike classification-based models, which often function as opaque \"black boxes\". Furthermore, developing a reliable classification model for atelectasis grading would require a much larger dataset to capture the wide variability in radiographic appearances, particularly subtle findings like linear opacities in grade 1—despite the positional invariance of CNNs. Hence, this segmentation-based method offers a more interpretable means of grading atelectasis and, by focusing solely on aerated volume, provides a broadly applicable solution across clinical contexts that remains effective even with limited training data. This study offers several key strengths. Firstly, it leverages a large real-world clinical cohort for automated quantification of atelectasis, using clinically interpretable AI-derived volumetry. By validating model outputs against physician-assigned atelectasis grades, the study exemplifies clinically applied AI, in contrast to much of AI research in healthcare that is decoupled from real-world outcomes. Unlike black-box classification models, our pipeline generates outputs that can be reviewed and validated by clinicians/technicians in real time, exemplifying verifiable AI. Its utility is further enhanced by the availability of open-source scripts for verifying segmentation accuracy and allowing integration into the backend of most software. A major technical strength is the use of nnU-Net—a widely validated segmentation framework in biomedical imaging—requiring relatively lesser expertise to deploy. Additionally, the volume metrics enable atelectasis to be measured as a continuous variable, preserving more information than categorical grading allowing for more nuanced statistical modeling. Moreover, continuous outcomes also provide greater statistical power and enable smaller sample sizes in clinical studies, supporting their use as efficient and meaningful endpoints in clinical trials. This work offers several potential avenues for both research and clinical application. From a research perspective, we have already applied this framework in a retrospective study (under review) assessing whether surgical fixation (“pexy”) of the RML reduces postoperative RML atelectasis and torsion 33 . Beyond this, the model outputs of lobar volumes/atelectasis could be used as quantitative endpoints in clinical trials focused on ventilation strategies and other interventions to prevent atelectasis, or on respiratory biomechanics across anesthesiology, critical care, and thoracic surgery. Clinically, the lobe segmentation model could be integrated into workflows such as preoperative planning of lung surgery and tumor localization through lobar mapping along with other models of tumor segmentation. It may also serve as an objective tool for postoperative assessment of atelectasis in research or clinical applications. These applications, however, warrant validation in prospective, context-specific studies. Despite the strengths of this study, several limitations must be acknowledged. While the preoperative model demonstrated strong generalizability, including external validation, caution is warranted in patients with chest wall or vertebral anomalies, and at the RUL/RML junction, where fissure incompleteness is often observed. Additionally, errors (Figs. 6 and 7 ) may stem from inconsistencies in reference standard annotations, a persistent challenge in biomedical AI. The postoperative model lacks external validation due to the absence of publicly available datasets. Furthermore, the clinical outcome of atelectasis grading was derived from a single-center retrospective cohort, which may limit generalizability. Hence until larger, more diverse datasets become available, applying automated lobe volumes for atelectasis grading may require visual confirmation. Special consideration is needed in cases with pleural pathology or external compression, which can globally affect lobar volumes and misrepresent aeration and relative lobar volumes. Importantly, the current pipeline is designed for longitudinal comparison and cannot interpret a single CT scan in isolation. Lastly, although we demonstrated the utility of continuous volume metrics as a surrogate for atelectasis, we did not define specific thresholds for grading atelectasis—an area that warrants further investigation and validation in future studies with clinical measurements like arterial blood gases, to establish meaningful thresholds for clinical classification. Conclusion We developed and validated a deep learning–based volumetric pipeline to quantify postoperative atelectasis after right upper lobectomy. Using automated lobar segmentation and validation against physician-assigned grades of atelectasis, we show that right middle lobe volume loss serves as an objective and reproducible surrogate for differentiating clinical atelectasis severity. These findings support the broader applicability of deep learning–driven lobar segmentation and volumetry as a reliable method for grading postoperative atelectasis in both clinical and research settings, and potentially for any atelectasis across other types of applications in pulmonary medicine and thoracic surgery. Declarations Code and data availability statement : The code and the final preoperative and postoperative model checkpoints are available on Github [https://github.com/Devanish31/atelectasis] and Figshare [10.6084/m9.figshare.29877509], respectively. IRB: Approved by the Stanford University School of Medicine IRB (IRB-70048). Author contributions : DNK – Conceptualization, Data curation, Formal analysis, Methodology, Supervision, Writing – original draft, Writing – review and editing; GMF - Conceptualization, Methodology, Writing – review and editing; NL - Conceptualization, Methodology, Writing – review and editing; LLT - Conceptualization, Methodology, Writing – review and editing; NSL - Conceptualization, Methodology, Writing – review and editing; IAE - Conceptualization, Methodology, Writing – review and editing; DZL - Conceptualization, Methodology, Writing – review and editing; LMB – Conceptualization, Methodology, Writing – review and editing; MFB - Conceptualization, Methodology, Writing – review and editing; HHG - Conceptualization, Methodology, Writing – review and editing; CPL - Conceptualization, Methodology, Writing – review and editing; JBS - Conceptualization, Methodology, Supervision, Writing – original draft, Writing – review and editing. All authors read and approved of the final manuscript. Declaration of generative AI and AI-assisted technologies : During the preparation of this work, the author(s) utilized ChatGPT to assist with rephrasing and refining the writing in the manuscript. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the published article. Competing interests : Joseph Shrager: Consulting - Becton Dickinson; Lungpacer; Leah M. Backhus, advisory panel member with Johnson and Johnson, AstraZeneca, Genentech/Roche, and Bristol Myers Squibb; Natalie S. Lui, consulting with Intuitive Surgical Inc. and Centese, grants from Intuitive Foundation; Other authors have nothing to declare. Curt Langlotz - This research is supported in part by the Medical Imaging and Data Resource Center; grant (75N92020D00021) from National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health and through The Advanced Research Projects Agency for Health (ARPA-H). Dr. Langlotz has received consulting fees from Sixth Street; honoraria for lectures and support for travel from Singapore Ministry of Health, McKinsey, and Philips; patent pending with GE HealthCare; former president of RSNA; on board of directors and shareholder of Bunkerhill Health (since March 31, 2019); on board of directors Sirona Medical (since February 2025) option holder of Whiterabbit.ai (since October 1, 2017); advisor and option holder of Galileo CDS (since May 1, 2019); advisor and option holder of Sirona Medical (since July 6, 2020); advisor and option holder of ADRA.ai (since September 17, 2020); advisor and option holder of Cognita (since November 13, 2024); Advisor and shareholder, TurboRadiology (since January 17, 2025); and gifts to institution, department, and/or research center from BunkerHill Health, Carestream, CARPL.ai, Clairity, GE HealthCare, Google Cloud, IBM, Kheiron, Lambda, Lunit, Microsoft, Nightingale Open Science, Philips, Siemens Healthineers, Stability.ai, Subtle Medical, VinBrain, Visiana, Whiterabbit.ai, Lowenstein Foundation, and Gordon and Betty Moore Foundation. Funding : None Acknowledgements : Some of the computing for this project was performed on the Sherlock cluster. We would like to thank Stanford University and the Stanford Research Computing Center for providing computational resources and support that contributed to these research results. This research used data/services provided by STARR (STAnford medicine Research data Repository), a clinical data warehouse containing patients' health information of the Stanford hospitals made possible by Stanford School of Medicine Research Office and the Stanford Clinical and Translational Science Award Number UL1TR003142 from the National Center for Advancing Translational Sciences. Data from the Stanford customized version of The Society of Thoracic Surgery Database was also used for creating this study cohort. References Zeng C, Lagier D, Lee JW, Vidal Melo MF. Perioperative Pulmonary Atelectasis: Part I. Biology and Mechanisms. Anesthesiology . 2022;136(1):181-205. Tanner TG, Colvin MO. Pulmonary Complications of Cardiac Surgery. Lung . 2020;198(6):889-896. Zhao Y, Zheng R, Xiang W, Ning D, Li Z. Systematic review and meta-analysis on perioperative intervention to prevent postoperative atelectasis complications after thoracic surgery. Ann Palliat Med . 2021;10(10):107260734-107210734. Ko E, Yoo KY, Lim CH, Jun S, Lee K, Kim YH. Is atelectasis related to the development of postoperative pneumonia? a retrospective single center study. BMC Anesthesiol . 2023;23:77. Newman B, Krane EJ, Gawande R, Holmes TH, Robinson TE. Chest CT in children: anesthesia and atelectasis. Pediatr Radiol . 2014;44(2):164-172. Khan A, Bashour S, Sabath B, et al. Severity of Atelectasis during Bronchoscopy: Descriptions of a New Grading System (AtelectasisSeverity Scoring System-\"ASSESS\") and At-Risk-Lung Zones. Diagn Basel Switz . 2024;14(2):197. McGuinness S, Parke RL. A New System For Assessing Atelectasis On Chest X-Ray After Sternotomy For Cardiac Surgery. In: A48. CLINICAL RESEARCH . American Thoracic Society International Conference Abstracts. American Thoracic Society; 2011:A1679-A1679. Auctores. A New System for Outcome Measure of Atelectasis in Studies of Cardiac Surgery. Auctores. Accessed June 27, 2025. https://www.auctoresonline.org/article/a-new-system-for-outcome-measure-of-atelectasis-in-studies-of-cardiac-surgery Sargent MA, McEachern AM, Jamieson DH, Kahwaji R. Atelectasis on pediatric chest CT: comparison of sedation techniques. Pediatr Radiol . 1999;29(7):509-513. Tenling A, Hachenberg T, Tydén H, Wegenius G, Hedenstierna G. Atelectasis and gas exchange after cardiac surgery. Anesthesiology . 1998;89(2):371-378. Reber A, Nylund U, Hedenstierna G. Position and shape of the diaphragm: implications for atelectasis formation. Anaesthesia . 1998;53(11):1054-1061. Warner DO, Warner MA, Ritman EL. Atelectasis and chest wall shape during halothane anesthesia. Anesthesiology . 1996;85(1):49-59. Pieters ALP, van der Veer T, Meerburg JJ, et al. Structural Lung Disease and Clinical Phenotype in Bronchiectasis Patients: The EMBARC CT Study. Am J Respir Crit Care Med . 2024;210(1):87-96. Rosenow T, Oudraad MCJ, Murray CP, et al. PRAGMA-CF. A Quantitative Structural Lung Disease Computed Tomography Outcome in Young Children with Cystic Fibrosis. Am J Respir Crit Care Med . 2015;191(10):1158-1165. Levine BE, Johnson RP. Effects of atelectasis on pulmonary surfactant and quasi-static lung mechanics. J Appl Physiol . 1965;20(5):859-864. Noll E, Ohana M, Hengen M, et al. Validation of MRI for Volumetric Quantification of Atelectasis in the Perioperative Period: An Experimental Study in Swine. Front Physiol . 2019;10:695. David M, Karmrodt J, Bletz C, et al. Analysis of atelectasis, ventilated, and hyperinflated lung during mechanical ventilation by dynamic CT. Chest . 2005;128(5):3757-3770. Mizukami Y, Takahashi Y, Maki R, Adachi H. Risk factors for atelectasis of the middle lobe after right upper lobectomy: preoperative bronchial diameter and stapling of the fissure. J Thorac Dis . 2021;13(10):5649-5657. Tejani AS, Klontzas ME, Gatti AA, et al. Checklist for Artificial Intelligence in Medical Imaging (CLAIM): 2024 Update. Radiol Artif Intell . 2024;6(4):e240300. D Slicer image computing platform. 3D Slicer. Accessed July 26, 2025. https://slicer.org/ Amini E, Klein R. Lung Lobe Segmentation and Quantification Dataset. Published online July 8, 2024. Accessed June 14, 2025. https://zenodo.org/records/12690803 Isensee F, Petersen J, Klein A, et al. nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation. arXiv:1809.10486. Isensee F, Wald T, Ulrich C, et al. nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation. arXiv:2404.09556. LOLA11 - Grand Challenge. grand-challenge.org. Accessed August 21, 2025. https://lola11.grand-challenge.org/Details/ Bragman FJS, McClelland JR, Jacob J, Hurst JR, Hawkes DJ. Pulmonary Lobe Segmentation with Probabilistic Segmentation of the Fissures and a Groupwise Fissure Prior. IEEE Trans Med Imaging . 2017;36(8):1650-1663. Imran AAZ, Hatamizadeh ,Ali, Ananth ,Shilpa P., Ding ,Xiaowei, Tajbakhsh ,Nima, and Terzopoulos D. Fast and automatic segmentation of pulmonary lobes from chest CT using a progressive dense V-network. Comput Methods Biomech Biomed Eng Imaging Vis . 2020;8(5):509-518. Xie W, Jacobs C, Charbonnier JP, van Ginneken B. Relational Modeling for Robust and Efficient Pulmonary Lobe Segmentation in CT Scans. IEEE Trans Med Imaging . 2020;39(8):2664-2675. Peng Y, Zhang J. Lung lobe segmentation in computed tomography images based on multi-feature fusion and ensemble learning framework. Int J Imaging Syst Technol . 2023;33(6):2088-2099. Pang H, Wu Y, Qi S, et al. A fully automatic segmentation pipeline of pulmonary lobes before and after lobectomy from computed tomography images. Comput Biol Med . 2022;147:105792. Malik H, Anees T, Al-Shamaylehs AS, Alharthi SZ, Khalil W, Akhunzada A. Deep Learning-Based Classification of Chest Diseases Using X-rays, CT Scans, and Cough Sound Images. Diagn Basel Switz . 2023;13(17):2772. Malik H, Anees T. Multi-modal deep learning methods for classification of chest diseases using different medical imaging and cough sounds. PloS One . 2024;19(3):e0296352. Rajpurkar P, Irvin J, Ball RL, et al. Deep learning for chest radiograph diagnosis: A retrospective comparison of the CheXNeXt algorithm to practicing radiologists. PLOS Med . 2018;15(11):e1002686. Kamtam DN, Liou DZ, Berry MF, et al. Does fixation of the right middle lobe/lingula during right upper lobectomy/left apical trisegmentectomy reduce rates of postoperative torsion/atelectasis? https://doi.org/10.1101/2025.07.28.25332308. Doel T, Matin TN, Gleeson FV, Gavaghan DJ, Grau V. Pulmonary lobe segmentation from CT images using fissureness, airways, vessels and multilevel B-splines. In: 2012 9th IEEE International Symposium on Biomedical Imaging (ISBI) . ; 2012:1491-1494. D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation | SpringerLink. Accessed June 26, 2025. https://link.springer.com/chapter/10.1007/978-3-319-46723-8_49 Park J, Yun J, Kim N, et al. Fully Automated Lung Lobe Segmentation in Volumetric Chest CT with 3D U-Net: Validation with Intra- and Extra-Datasets. J Digit Imaging . 2020;33(1):221-230. Lassen-Schmidt B, Hering A, Krass S, Meine H. Automatic segmentation of the pulmonary lobes with a 3D u-net and optimized loss function. arXiv:2006.00083. Ferreira FT, Sousa P, Galdran A, Sousa MR, Campilho A. End-to-End Supervised Lung Lobe Segmentation. In: 2018 International Joint Conference on Neural Networks (IJCNN) . ; 2018:1-8. Gerard SE, Reinhardt JM. Pulmonary Lobe Segmentation Using A Sequence of Convolutional Neural Networks For Marginal Learning. In: 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019) . ; 2019:1207-1211. Zheng S, Nie W, Pan L, et al. A dual-attention V-network for pulmonary lobe segmentation in CT scans. IET Image Process . 2021;15(8):1644-1654. Zhang Z, Ren J, Tao X, et al. Automatic segmentation of pulmonary lobes on low-dose computed tomography using deep learning. Ann Transl Med . 2021;9(4):291. Bao N, Yuan Y, Luo Q, Li Q, Zhang LB. Edge-enhancement cascaded network for lung lobe segmentation based on CT images. Front Phys . 2023;11. Nomura GRO, Luong AT, Prakash A, et al. TriSwinUNETR lobe segmentation model for computing DIR-free CT-ventilation. Front Oncol . 2025;15:1475133. van Rikxoort EM, Prokop M, de Hoop B, Viergever MA, Pluim JPW, van Ginneken B. Automatic segmentation of pulmonary lobes robust against incomplete fissures. IEEE Trans Med Imaging . 2010;29(6):1286-1296. Lassen B, van Rikxoort EM, Schmidt M, Kerkstra S, van Ginneken B, Kuhnigk JM. Automatic segmentation of the pulmonary lobes from chest CT scans based on fissures, vessels, and bronchi. IEEE Trans Med Imaging . 2013;32(2):210-222. Additional Declarations Competing interest reported. Joseph Shrager: Consulting - Becton Dickinson; Lungpacer; Leah M. Backhus, advisory panel member with Johnson and Johnson, AstraZeneca, Genentech/Roche, and Bristol Myers Squibb; Natalie S. Lui, consulting with Intuitive Surgical Inc. and Centese, grants from Intuitive Foundation; Other authors have nothing to declare. Curt Langlotz - This research is supported in part by the Medical Imaging and Data Resource Center; grant (75N92020D00021) from National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health and through The Advanced Research Projects Agency for Health (ARPA-H). Dr. Langlotz has received consulting fees from Sixth Street; honoraria for lectures and support for travel from Singapore Ministry of Health, McKinsey, and Philips; patent pending with GE HealthCare; former president of RSNA; on board of directors and shareholder of Bunkerhill Health (since March 31, 2019); on board of directors Sirona Medical (since February 2025) option holder of Whiterabbit.ai (since October 1, 2017); advisor and option holder of Galileo CDS (since May 1, 2019); advisor and option holder of Sirona Medical (since July 6, 2020); advisor and option holder of ADRA.ai (since September 17, 2020); advisor and option holder of Cognita (since November 13, 2024); Advisor and shareholder, TurboRadiology (since January 17, 2025); and gifts to institution, department, and/or research center from BunkerHill Health, Carestream, CARPL.ai, Clairity, GE HealthCare, Google Cloud, IBM, Kheiron, Lambda, Lunit, Microsoft, Nightingale Open Science, Philips, Siemens Healthineers, Stability.ai, Subtle Medical, VinBrain, Visiana, Whiterabbit.ai, Lowenstein Foundation, and Gordon and Betty Moore Foundation. Supplementary Files SupplementaryFigures.docx Cite Share Download PDF Status: Published Journal Publication published 30 Apr, 2026 Read the published version in npj Digital Medicine → Version 1 posted Editorial decision: Revision requested 03 Nov, 2025 Reviewers agreed at journal 03 Nov, 2025 Reviewers agreed at journal 02 Nov, 2025 Reviews received at journal 01 Nov, 2025 Reviewers agreed at journal 31 Oct, 2025 Reviews received at journal 31 Oct, 2025 Reviewers agreed at journal 31 Oct, 2025 Reviews received at journal 20 Oct, 2025 Reviewers agreed at journal 20 Oct, 2025 Reviewers agreed at journal 18 Oct, 2025 Reviewers agreed at journal 16 Oct, 2025 Reviewers invited by journal 16 Oct, 2025 Editor assigned by journal 13 Oct, 2025 Submission checks completed at journal 13 Oct, 2025 First submitted to journal 02 Oct, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {\"props\":{\"pageProps\":{\"initialData\":{\"identity\":\"rs-7768040\",\"acceptedTermsAndConditions\":true,\"allowDirectSubmit\":false,\"archivedVersions\":[],\"articleType\":\"Article\",\"associatedPublications\":[],\"authors\":[{\"id\":525633334,\"identity\":\"0b5801fc-0a38-4dd6-9c87-0f6ccbd266c7\",\"order_by\":0,\"name\":\"Devanish N. Kamtam\",\"email\":\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5klEQVRIie3NvarCMBTA8UCg07Gut+QljhQK0j5MQqDjteBeBFehawUfQrhw55SAXQqugQ4igpObIIIOfq0S6uaQ/3BIwvkRQlyub0w9BhLef96S18tPJxJM6P2QdiaEcFRdiV/PBscsy3/D9bo6ZjwZ+YpWLVhI0DQhK1GPIyMpK3k6nCtPxjaCJuUMUIl/QwnrnTWigohZyWYvL4C5+Cs0vQB/kP7JTgxd3X+hYkmkx14EPCsJGrmKAbUojYxi4CkG2guHCwvx62rawjUXRVHtWuAJ+vV0aw4W8ib62brL5XK53nQDUztNhYH2irkAAAAASUVORK5CYII=\",\"orcid\":\"\",\"institution\":\"Stanford University\",\"correspondingAuthor\":true,\"prefix\":\"\",\"firstName\":\"Devanish\",\"middleName\":\"N.\",\"lastName\":\"Kamtam\",\"suffix\":\"\"},{\"id\":525633335,\"identity\":\"1215ea0a-64c1-46e5-a707-d17cf75413c5\",\"order_by\":1,\"name\":\"Guiseppe M. Facchi\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"University of Milan\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Guiseppe\",\"middleName\":\"M.\",\"lastName\":\"Facchi\",\"suffix\":\"\"},{\"id\":525633339,\"identity\":\"faebbb3c-e5b5-49e5-87a6-4feedc6731a5\",\"order_by\":2,\"name\":\"Nicole Lin\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Stanford University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Nicole\",\"middleName\":\"\",\"lastName\":\"Lin\",\"suffix\":\"\"},{\"id\":525633341,\"identity\":\"4bd410e5-5d07-4225-a9e8-a3e78f16bcde\",\"order_by\":3,\"name\":\"Lillian L. Tsai\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Stanford University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Lillian\",\"middleName\":\"L.\",\"lastName\":\"Tsai\",\"suffix\":\"\"},{\"id\":525633343,\"identity\":\"cf49e16e-6dbb-4e0b-aaf8-437014eb3420\",\"order_by\":4,\"name\":\"Natalie S. Lui\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Stanford University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Natalie\",\"middleName\":\"S.\",\"lastName\":\"Lui\",\"suffix\":\"\"},{\"id\":525633345,\"identity\":\"0b9dd913-3e4d-4a56-a154-40d77c585ce0\",\"order_by\":5,\"name\":\"Irmina A. Elliott\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Stanford University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Irmina\",\"middleName\":\"A.\",\"lastName\":\"Elliott\",\"suffix\":\"\"},{\"id\":525633347,\"identity\":\"3d71a6bd-bc78-4c7c-96b3-923e3b0a4de2\",\"order_by\":6,\"name\":\"Doug Z. Liou\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Stanford University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Doug\",\"middleName\":\"Z.\",\"lastName\":\"Liou\",\"suffix\":\"\"},{\"id\":525633348,\"identity\":\"1a3368b3-b1a7-4698-b4ea-9f49985dd0ad\",\"order_by\":7,\"name\":\"Leah M. Backhus\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Stanford University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Leah\",\"middleName\":\"M.\",\"lastName\":\"Backhus\",\"suffix\":\"\"},{\"id\":525633350,\"identity\":\"6fe967cd-cef3-4334-8827-6836f4a76cd0\",\"order_by\":8,\"name\":\"Mark F. Berry\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Stanford University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Mark\",\"middleName\":\"F.\",\"lastName\":\"Berry\",\"suffix\":\"\"},{\"id\":525633351,\"identity\":\"f4f8fd60-aee6-4431-9856-42b30630b377\",\"order_by\":9,\"name\":\"Henry H. Guo\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Stanford University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Henry\",\"middleName\":\"H.\",\"lastName\":\"Guo\",\"suffix\":\"\"},{\"id\":525633355,\"identity\":\"b9d55de0-0d52-4eb9-b8ab-bbf922e95187\",\"order_by\":10,\"name\":\"Curtis P. Langlotz\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Stanford University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Curtis\",\"middleName\":\"P.\",\"lastName\":\"Langlotz\",\"suffix\":\"\"},{\"id\":525633356,\"identity\":\"27f8abd5-3a32-415e-b2c0-5eae26cab778\",\"order_by\":11,\"name\":\"Joseph B. Shrager\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Stanford University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Joseph\",\"middleName\":\"B.\",\"lastName\":\"Shrager\",\"suffix\":\"\"}],\"badges\":[],\"createdAt\":\"2025-10-02 16:38:14\",\"currentVersionCode\":1,\"declarations\":\"\",\"doi\":\"10.21203/rs.3.rs-7768040/v1\",\"doiUrl\":\"https://doi.org/10.21203/rs.3.rs-7768040/v1\",\"draftVersion\":[],\"editorialEvents\":[{\"content\":\"https://doi.org/10.1038/s41746-026-02683-6\",\"type\":\"published\",\"date\":\"2026-04-30T15:57:09+00:00\"}],\"editorialNote\":\"\",\"failedWorkflow\":false,\"files\":[{\"id\":93027685,\"identity\":\"4929a9ac-b1e8-4156-bd16-23b705fb32fe\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:42:21\",\"extension\":\"docx\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":4973189,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"FinalManuscript.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/10d4db32d1ead9f128e0ac06.docx\"},{\"id\":93026053,\"identity\":\"83aac711-e5ca-4c06-9809-2897ebb61ca5\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:26:21\",\"extension\":\"json\",\"order_by\":1,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":15346,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"25cf360adeb84eaaa99c411e73127286.json\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/02d30cde0618dbe2908bb7a4.json\"},{\"id\":93024710,\"identity\":\"e1a84c35-4dcb-4961-8432-aa5a774498c2\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"xml\",\"order_by\":2,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":163171,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"25cf360adeb84eaaa99c411e731272861enriched.xml\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/ca43f16d002fe9f8303a8c6a.xml\"},{\"id\":93024708,\"identity\":\"524a04b9-899e-4be7-b5f9-db4fa41e9f8d\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"jpeg\",\"order_by\":3,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":272910,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage1.jpeg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/cb1aff69f6f7470db57e7dae.jpeg\"},{\"id\":93024705,\"identity\":\"509f6b66-cb90-4e3b-986c-a3e56c89e519\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"jpeg\",\"order_by\":4,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":392452,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage10.jpeg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/26dd016f369a502ab64a97f4.jpeg\"},{\"id\":93026059,\"identity\":\"28505586-5746-4ecc-befb-7902e6b1c5f1\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:26:21\",\"extension\":\"jpeg\",\"order_by\":5,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":384378,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage2.jpeg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/b04f47cdc3c85ec18bc5b7ed.jpeg\"},{\"id\":93024713,\"identity\":\"935d38a6-01b6-44ca-a8bd-b7dfcdf17c47\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"jpeg\",\"order_by\":6,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":364826,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage3.jpeg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/b78e0c2840e0a2a9b42c3957.jpeg\"},{\"id\":93024722,\"identity\":\"4dc19a21-2930-4d80-b422-270273bd77fa\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"png\",\"order_by\":7,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":1377266,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage4.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/992da6dde0f558d3e489dea3.png\"},{\"id\":93024731,\"identity\":\"4308083c-b92d-4143-ad6a-6e640ee7187a\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:22\",\"extension\":\"png\",\"order_by\":8,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":812590,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage5.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/d9732cb8af816dc9ed280183.png\"},{\"id\":93026062,\"identity\":\"b2ce78c6-5651-43dc-a002-e4e9c4f18b3b\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:26:21\",\"extension\":\"jpeg\",\"order_by\":9,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":1185787,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage6.jpeg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/f8953b3ebc4d8858135eac7d.jpeg\"},{\"id\":93026781,\"identity\":\"7fa39837-e29c-4f6e-8209-660025c65342\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:34:21\",\"extension\":\"png\",\"order_by\":10,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":1253766,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage7.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/7f34f8b9a0de3c8ef5284d1d.png\"},{\"id\":93024715,\"identity\":\"c3b69564-b658-480c-b854-3311f2f811c0\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"png\",\"order_by\":11,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":50159,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage8.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/3e8c372a98d7b6486a4e8eef.png\"},{\"id\":93024726,\"identity\":\"41eb33de-a6eb-4179-99a5-007212baca0b\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"png\",\"order_by\":12,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":87680,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage9.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/fecb5c62caeb2f2287c635d1.png\"},{\"id\":93024720,\"identity\":\"82b80d06-9135-4151-9c0c-3c3ae2996e2b\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"png\",\"order_by\":13,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":65269,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage1.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/c2bd1a1e7eb0db13ecbfbbf2.png\"},{\"id\":93026061,\"identity\":\"f6164ccc-985c-4d95-b93b-634cb1746aed\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:26:21\",\"extension\":\"png\",\"order_by\":14,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":84837,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage10.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/ef217ee17ff86bb4ffad455b.png\"},{\"id\":93024718,\"identity\":\"d2c26415-fe8e-4e00-98f6-18f4434a87d1\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"png\",\"order_by\":15,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":89088,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage2.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/a6fcc67a59e990ddaeaaf254.png\"},{\"id\":93026064,\"identity\":\"196181f3-40a6-40e5-9404-05fc637e1066\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:26:21\",\"extension\":\"png\",\"order_by\":16,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":76116,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage3.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/4d770727353c1b41c27a7ab0.png\"},{\"id\":93024728,\"identity\":\"f5838c34-b534-4ac9-8c18-c87f631eddc3\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"png\",\"order_by\":17,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":145694,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage4.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/66923320326ea15f7538d212.png\"},{\"id\":93024725,\"identity\":\"e56e1ae0-c7c3-419f-999d-9fa252bd9e64\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"png\",\"order_by\":18,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":155694,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage5.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/90a0bb9c60d7eb7535bf6af1.png\"},{\"id\":93024719,\"identity\":\"793f1c6d-74e6-485d-b529-8a81c63deb86\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"png\",\"order_by\":19,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":231237,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage6.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/e81ee47b8dec610b43476030.png\"},{\"id\":93024727,\"identity\":\"619c3b46-3e04-4995-b3a3-8d7b7c2d9ade\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"png\",\"order_by\":20,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":139411,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage7.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/4a92642a813bedfc5dc4bdab.png\"},{\"id\":93024716,\"identity\":\"584c07a7-f786-4d66-bded-52af623a279c\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"png\",\"order_by\":21,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":18177,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage8.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/3fa8c1f954b2c2800eb9b605.png\"},{\"id\":93024724,\"identity\":\"301dc79e-c760-45b2-b713-565c9bab00c1\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"png\",\"order_by\":22,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":32245,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage9.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/f8589b8c8992a62128658da2.png\"},{\"id\":93024730,\"identity\":\"2e5f64cf-9d93-45a6-9e56-84ed86e9d5e6\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:22\",\"extension\":\"xml\",\"order_by\":23,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":162746,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"25cf360adeb84eaaa99c411e731272861structuring.xml\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/6a1c61366c4e84a1d7951ba1.xml\"},{\"id\":93024732,\"identity\":\"4981fdf5-cc03-4e53-b566-fca5f3a020f2\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:22\",\"extension\":\"html\",\"order_by\":24,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":178095,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"earlyproof.html\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/3aae11d5590868b5090654e9.html\"},{\"id\":93026054,\"identity\":\"0605e838-f43f-4e6d-a354-1f8490a70864\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:26:21\",\"extension\":\"jpg\",\"order_by\":1,\"title\":\"Figure 1\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":91829,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eOverview of the model development pipelines using preoperative (A) and postoperative (B) CT datasets.\\u003c/p\\u003e\\n\\u003cp\\u003e(A) Preoperative CT model development: A total of 122 scans from Ottawa (59 training + 22 test) and Stanford (34 training + 7 test) chest CTs were combined. The dataset was split into 93 training and 29 validation scans, used to train a 3D nnUNetv2 full-resolution model for 5000 epochs with Dice+Cross Entropy loss. The best-performing model checkpoint was evaluated on both the internal test set (n=29) and an external test set (LOLA11 dataset, n=55).\\u003c/p\\u003e\\n\\u003cp\\u003e(B) Postoperative CT model development: A separate cohort of 82 Stanford postoperative chest CTs (67 for training, 15 for testing) was used to train a second 3D nnUNetv2 model using the same settings.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"1.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/127c1c238867a93cf386f180.jpg\"},{\"id\":93026056,\"identity\":\"4702ca41-d14f-4dd3-9b4c-5548e7beb546\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:26:21\",\"extension\":\"jpg\",\"order_by\":2,\"title\":\"Figure 2\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":165855,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eTraining and validation loss curves along with pseudo-Dice similarity coefficient (DSC) across epochs for the preoperative (A) and postoperative (B) models.\\u003cbr\\u003e\\nIn both panels, the blue line represents training loss, red line represents validation loss, dotted green line represents raw pseudo-Dice, and solid green line represents its moving average.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"2.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/82644d8625bc4f9418c721f6.jpg\"},{\"id\":93024703,\"identity\":\"3314deaa-a50e-491a-8227-6f5595d0a220\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"jpg\",\"order_by\":3,\"title\":\"Figure 3\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":160532,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eBoxplots showing changes in right middle lobe (RML) and right lower lobe (RLL) volume (ΔRML, ΔRLL) and their normalized ratios with respect to right lung (RL) and total lung (TL) volumes across increasing grades of atelectasis. The top row (A-E) shows the study cohort stratified into five atelectasis grades (0–4: none, minimal, subsegmental, segmental, and lobar). The bottom row (F-J) shows pooled atelectasis grades (none, minimal-subsegmental, segmental-lobar). Asterisks indicate statistically significant pairwise comparisons (with Bonferroni correction).\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"3.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/85e100c7588d2052e21db889.jpg\"},{\"id\":93024701,\"identity\":\"41157b7a-100e-453b-a15f-a4aebca26cae\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"jpg\",\"order_by\":4,\"title\":\"Figure 4\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":132748,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eAxial CT slices showing preoperative (left column) and 6-month postoperative (right column) lobe segmentations in patients with increasing grades of atelectasis. The lobe in yellow in right column depicts the progressively increasing grades of atelectasis: minimal (grade 1) (E.), subsegmental (grade 2) (F.), segmental (grade 3) (G.), and lobar atelectasis (grade 4) (H.). Lobe labels (colors): RUL (green), RML (yellow), RLL (brown), LUL (blue), LLL (red), LL (pink).\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"4.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/7a03313cd7c76f887a7e8b10.jpg\"},{\"id\":93026055,\"identity\":\"038702b1-9068-4e92-a6cd-2e609a7f656a\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:26:21\",\"extension\":\"jpg\",\"order_by\":5,\"title\":\"Figure 5\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":75946,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e3D modelling of lobe segmentations of preoperative (left) and 6-month postoperative (right) CT scans in patients with increasing severity of postoperative atelectasis. The lobe in yellow in the right column depicts the progressively increasing grades of atelectasis: minimal (E.), subsegmental (F.), segmental (G.), and lobar atelectasis (H.). Lobe labels (colors): RUL (green), RML (yellow), RLL (brown), LUL (blue), LLL (red), LL (pink).\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"5.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/d0711ee1e43acbb39269de4c.jpg\"},{\"id\":93026057,\"identity\":\"283a7a0d-9b6b-4fb6-93ce-8f0960264da1\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:26:21\",\"extension\":\"jpg\",\"order_by\":6,\"title\":\"Figure 6\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":142081,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eRepresentative segmentation errors on training and internal test dataset vs. reference standard annotations (green - false positive prediction; red - false negative prediction).\\u003c/p\\u003e\\n\\u003cp\\u003e(A.) right upper lobe/middle lobe junction region.\\u003cbr\\u003e\\n(B.) right middle lobe/lower lobe junction region, particularly at the hilum.\\u003cbr\\u003e\\n(C.) bullae of COPD chest CT.\\u003cbr\\u003e\\n(D.) consolidated lung or adjacent structures.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"6.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/66080c79ad93d858f6812f17.jpg\"},{\"id\":93024711,\"identity\":\"d73f41d7-6e7d-4c5a-aa87-b373264a5e75\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"jpg\",\"order_by\":7,\"title\":\"Figure 7\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":116439,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eExamples of the challenging cases in the external validation dataset\\u003cstrong\\u003e \\u003c/strong\\u003e(LOLA11)\\u003cstrong\\u003e \\u003c/strong\\u003e(A)\\u003cstrong\\u003e \\u003c/strong\\u003eScoliosis (B) Severe COPD (Emphysema) (C) Chestwall deformity (D) Right-sided complete pleural space opacification. Lobe labels (colors): RUL (green), RML (yellow), RLL (brown), LUL (blue), LLL (red), LL (pink).\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"7.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/b840b4bce66a9916b7cba3bc.jpg\"},{\"id\":108437608,\"identity\":\"1ec1b25b-efe2-4ffe-8ef8-9c3cbf8972cd\",\"added_by\":\"auto\",\"created_at\":\"2026-05-04 16:00:18\",\"extension\":\"pdf\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"manuscript-pdf\",\"size\":1439223,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"manuscript.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/819b756d-da33-4756-835e-c92249210747.pdf\"},{\"id\":93024699,\"identity\":\"037d42a9-dc8f-41c1-b76b-8716fd87764d\",\"added_by\":\"auto\",\"created_at\":\"2025-10-08 09:18:21\",\"extension\":\"docx\",\"order_by\":1,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":252939,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryFigures.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-7768040/v1/d537effa9eb0ba288c7c6940.docx\"}],\"financialInterests\":\"Competing interest reported. Joseph Shrager: Consulting - Becton Dickinson; Lungpacer; Leah M. Backhus, advisory panel member with Johnson and Johnson, AstraZeneca, Genentech/Roche, and Bristol Myers Squibb; Natalie S. Lui, consulting with Intuitive Surgical Inc. and Centese, grants from Intuitive Foundation; Other authors have nothing to declare. Curt Langlotz - This research is supported in part by the Medical Imaging and Data Resource Center; grant (75N92020D00021) from National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health and through The Advanced Research Projects Agency for Health (ARPA-H). Dr. Langlotz has received consulting fees from Sixth Street; honoraria for lectures and support for travel from Singapore Ministry of Health, McKinsey, and Philips; patent pending with GE HealthCare; former president of RSNA; on board of directors and shareholder of Bunkerhill Health (since March 31, 2019); on board of directors Sirona Medical (since February 2025) option holder of Whiterabbit.ai (since October 1, 2017); advisor and option holder of Galileo CDS (since May 1, 2019); advisor and option holder of Sirona Medical (since July 6, 2020); advisor and option holder of ADRA.ai (since September 17, 2020); advisor and option holder of Cognita (since November 13, 2024); Advisor and shareholder, TurboRadiology (since January 17, 2025); and gifts to institution, department, and/or research center from BunkerHill Health, Carestream, CARPL.ai, Clairity, GE HealthCare, Google Cloud, IBM, Kheiron, Lambda, Lunit, Microsoft, Nightingale Open Science, Philips, Siemens Healthineers, Stability.ai, Subtle Medical, VinBrain, Visiana, Whiterabbit.ai, Lowenstein Foundation, and Gordon and Betty Moore Foundation.\",\"formattedTitle\":\"Deep-learning based quantitative evaluation of postoperative atelectasis following right upper lobectomy\",\"fulltext\":[{\"header\":\"Introduction\",\"content\":\"\\u003cp\\u003eAtelectasis \\u0026ndash; partial or complete collapse of a lung lobe resulting from alveolar deflation or inadequate expansion \\u0026ndash; is one of the most common pulmonary complications following surgery\\u003csup\\u003e\\u003cspan citationid=\\\"CR1\\\" class=\\\"CitationRef\\\"\\u003e1\\u003c/span\\u003e\\u003c/sup\\u003e. It is particularly common following cardiothoracic surgery, with rates of 30\\u0026ndash;72%\\u003csup\\u003e2,3\\u003c/sup\\u003e, versus 3.2% following non-cardiothoracic surgery\\u003csup\\u003e\\u003cspan citationid=\\\"CR4\\\" class=\\\"CitationRef\\\"\\u003e4\\u003c/span\\u003e\\u003c/sup\\u003e. Atelectasis can be classified into obstructive (resorptive), compressive, and adhesive types, based on underlying mechanisms such as internal airway obstruction, external compression, or surfactant dysfunction. Collapse of lung lobes reduces the alveolar surface area available for gas exchange, leading to impaired oxygenation, hypoxemia, and, in severe cases, respiratory failure. Further, completely atelectatic lung lobes and segments have a propensity to become infected.\\u003c/p\\u003e\\u003cp\\u003eWhile early postoperative atelectasis is often self-limited, complete lobar collapse and ongoing, substantial sublobar collapse can lead to significant morbidity, especially following lung resections where the remaining lung must compensate for the loss of resected parenchyma. Postoperative atelectasis is typically identified through radiographic imaging, with chest X-rays for preliminary assessment, followed by CT scans for diagnostic confirmation and more detailed assessment of the severity of atelectasis.\\u003c/p\\u003e\\u003cp\\u003eTo facilitate consistent evaluation of atelectasis \\u0026mdash;albeit largely in non-postoperative scenarios\\u0026mdash;several studies have proposed standardized grading criteria\\u003csup\\u003e\\u003cspan citationid=\\\"CR5\\\" class=\\\"CitationRef\\\"\\u003e5\\u003c/span\\u003e\\u003c/sup\\u003e, some of which have also been validated against clinical outcomes\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR7\\\" citationid=\\\"CR6\\\" class=\\\"CitationRef\\\"\\u003e6\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e\\u003c/sup\\u003e. Early objective approaches to grading atelectasis primarily relied on Hounsfield Unit (HU) measurements\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR10 CR11\\\" citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR12\\\" class=\\\"CitationRef\\\"\\u003e12\\u003c/span\\u003e\\u003c/sup\\u003e. CT slice intensity values between \\u0026minus;\\u0026thinsp;500 and +\\u0026thinsp;100 HU (vs. \\u0026lt; -500 HU as aerated lung) have been used to identify poorly aerated regions reflecting atelectasis\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR10 CR11\\\" citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR12\\\" class=\\\"CitationRef\\\"\\u003e12\\u003c/span\\u003e\\u003c/sup\\u003e, and these metrics demonstrated good correlation with clinical outcomes. While effective within the lung parenchyma, this approach becomes less reliable when atelectasis extends to the lung periphery, as the collapsed lung border may be difficult to distinguish from adjacent mediastinal structures on CT. And in complete lobar collapse, the proportion of collapsed voxels may be outweighed by the volume of the remaining aerated lung, leading to underestimation of atelectasis severity. Moreover, despite HU standardization (\\u0026ndash;1000 HU - air, 0 HU - water), the HU thresholds may differ by scanner type/ imaging acquisition protocols (Supplementary Fig.\\u0026nbsp;1).\\u003c/p\\u003e\\u003cp\\u003eMore recently, other methods, such as the BEST-CT\\u003csup\\u003e13\\u003c/sup\\u003e, have been used to quantify atelectasis as a percentage of total lung volume. This method divides axial CT slices into grids and classifies each segment into one of ten status categories, including atelectasis, the percentage of which is then calculated across the entire scan. While this methodology has been validated in two clinical contexts\\u0026mdash;cystic fibrosis\\u003csup\\u003e\\u003cspan citationid=\\\"CR14\\\" class=\\\"CitationRef\\\"\\u003e14\\u003c/span\\u003e\\u003c/sup\\u003e and bronchiectasis\\u003csup\\u003e\\u003cspan citationid=\\\"CR13\\\" class=\\\"CitationRef\\\"\\u003e13\\u003c/span\\u003e\\u003c/sup\\u003e\\u0026mdash; the grading remains entirely manual and time-intensive. Recently, less time-intensive scoring systems like the ASSESS criteria have been introduced in post-interventional settings such as following bronchoscopy under general anesthesia, grading atelectasis by its extent on CT from the posterior chest wall to the anterior vertebral border. Their applicability, however, is limited to dependent dorso-caudal lung regions.\\u003c/p\\u003e\\u003cp\\u003eOverall, existing methods of grading atelectasis remain constrained by their manual, condition-specific, subjective nature, with significant interrater variability. This limits their generalizability, scalability, and reproducibility. Deep learning may enable more robust and automated grading of atelectasis that is reproducible across diverse clinical contexts, imaging protocols, and healthcare systems\\u0026mdash;provided such variability is adequately represented in the training data. Atelectasis can be graded either by segmenting and quantifying the reduction in volume of the collapsed regions of the lung\\u003csup\\u003e\\u003cspan citationid=\\\"CR5\\\" class=\\\"CitationRef\\\"\\u003e5\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e\\u003c/sup\\u003e or, alternatively, by segmenting and quantifying volume of the aerated regions that represent the remaining functional lung. Prior studies\\u003csup\\u003e\\u003cspan citationid=\\\"CR15\\\" class=\\\"CitationRef\\\"\\u003e15\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR16\\\" class=\\\"CitationRef\\\"\\u003e16\\u003c/span\\u003e\\u003c/sup\\u003e have confirmed that quantitative volumetric analysis of lung aeration\\u0026mdash;whether by CT or MRI\\u003csup\\u003e\\u003cspan citationid=\\\"CR16\\\" class=\\\"CitationRef\\\"\\u003e16\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR17\\\" class=\\\"CitationRef\\\"\\u003e17\\u003c/span\\u003e\\u003c/sup\\u003e\\u0026mdash;provides objective metrics that can serve as reliable proxies for assessing the degree of atelectasis. Hence, we aimed to use automated volumetric analysis of aerated lung regions to quantify and grade the severity of atelectasis.\\u003c/p\\u003e\\u003cp\\u003eGiven its higher incidence following thoracic surgery, we evaluated atelectasis following pulmonary lobectomy, using right upper lobectomy (RULobecttomy)\\u0026mdash;the most frequently performed lobectomy\\u0026mdash;as a representative model. This was also an ideal model because chronic right middle lobe (RML) atelectasis, often referred to as right middle lobe syndrome, is a recognized long-term complication after right upper lobectomy\\u003csup\\u003e\\u003cspan citationid=\\\"CR18\\\" class=\\\"CitationRef\\\"\\u003e18\\u003c/span\\u003e\\u003c/sup\\u003e. Given this known risk, we specifically targeted the quantification of middle lobe volume loss and its correlation with physician-assigned atelectasis grades.\\u003c/p\\u003e\\u003cp\\u003eWe leveraged paired preoperative and six-month-postoperative CT scans from patients who had undergone RULobectomy to develop and validate an automated, deep learning\\u0026ndash;based volumetric pipeline to quantify/grade atelectasis. Our aim was to establish a reproducible and clinically interpretable framework for quantifying lobar volume changes and grading postoperative atelectasis and to evaluate its clinical relevance by comparing volumetry-derived atelectasis metrics with physician-assigned grades in a large postoperative cohort. This approach may offer a scalable tool for both retrospective research and prospective clinical risk stratification, with potential applicability to other lobectomies and clinical contexts in which atelectasis occurs.\\u003c/p\\u003e\"},{\"header\":\"Methods\",\"content\":\"\\u003cdiv id=\\\"Sec3\\\" class=\\\"Section2\\\"\\u003e\\u003ch2\\u003eStudy design\\u003c/h2\\u003e\\u003cp\\u003eThis study was a single-center retrospective analysis. The Checklist for Artificial Intelligence in Medical Imaging (CLAIM-2024) was used in study design and implementation\\u003csup\\u003e\\u003cspan citationid=\\\"CR19\\\" class=\\\"CitationRef\\\"\\u003e19\\u003c/span\\u003e\\u003c/sup\\u003e. The CLAIM checklist is modeled after the Standards for Reporting of Diagnostic Accuracy Studies (STARD) guidelines and is part of the EQUATOR network, ensuring best practices in reporting and facilitating the translation of AI into clinical practice. The study was approved by the Stanford University School of Medicine Institutional Review Board (IRB), which waived the requirement for informed consent as the study involved only retrospective chart review (IRB-70048). All procedures were carried out in accordance with the Declaration of Helsinki ethical standards.\\u003c/p\\u003e\\u003c/div\\u003e\\n\\u003ch3\\u003ePatient cohort and clinical data\\u003c/h3\\u003e\\n\\u003cp\\u003eWe retrospectively reviewed all consecutive patients who underwent a right upper lobectomy in the Division of Thoracic Surgery at Stanford University Hospital from January 2008 through December 2023. Patients were identified using Stanford\\u0026rsquo;s customized version of the Society of Thoracic Surgery (STS) General Thoracic Database and the STAnford Research Repository (STARR) clinical data warehouse. Patients without a preoperative CT scan within 6 months prior to surgery or a postoperative CT scan performed at approximately 6 months (range: 3\\u0026ndash;9 months) after surgery were excluded. Patients with scans of axial slice thickness\\u0026thinsp;\\u0026gt;\\u0026thinsp;3 mm were also excluded. Baseline demographic and clinical data were extracted from the databases and the electronic medical record. Follow-up data were collected for up to 6 months post-operatively.\\u003c/p\\u003e\\n\\u003ch3\\u003eOutcomes\\u003c/h3\\u003e\\n\\u003cp\\u003eAtelectasis of the RML was graded using a 5-point scale: 0\\u0026thinsp;=\\u0026thinsp;none, 1\\u0026thinsp;=\\u0026thinsp;minimal/linear, 2\\u0026thinsp;=\\u0026thinsp;subsegmental, 3\\u0026thinsp;=\\u0026thinsp;segmental, and 4\\u0026thinsp;=\\u0026thinsp;near-total or total lobar collapse. CT scans were reviewed using a lung window setting (window width: 1500 HU; window level: \\u0026minus;\\u0026thinsp;600 HU). Image evaluation was performed using 3D Slicer (version 8.5.1; MIT, Massachusetts, USA)\\u003csup\\u003e\\u003cspan citationid=\\\"CR20\\\" class=\\\"CitationRef\\\"\\u003e20\\u003c/span\\u003e\\u003c/sup\\u003e. Two board-certified surgeons independently reviewed and graded all scans, resolving discrepancies through consensus until full agreement was achieved.\\u003c/p\\u003e\\n\\u003ch3\\u003eImaging acquisition and radiological data\\u003c/h3\\u003e\\n\\u003cp\\u003ePreoperative and postoperative diagnostic chest CT scans were acquired using multidetector CT scanners from multiple vendors, most commonly Siemens (Siemens Healthineers, Erlangen, Germany) and GE Medical Systems (GE Healthcare, Chicago, IL, USA), occasionally with Toshiba (Toshiba Medical Systems Corporation, Otawara, Japan) and Philips (Philips Healthcare, Best, The Netherlands). Scans were performed during end-inspiratory breath-hold when tolerated. A wide variety of reconstruction kernels was used across the cohort, with the most frequent being STANDARD, T20f, Tr20f, and T20s. Axial images were reconstructed with slice thicknesses ranging from 0.3 mm to 3 mm.\\u003c/p\\u003e\\u003cp\\u003eCT acquisition protocols varied, with tube voltages ranging from 80 to 130 kVp (median [IQR]: preoperative \\u0026ndash; 120 [120\\u0026ndash;120]; postoperative \\u0026ndash; 120 [80\\u0026ndash;120]) and tube currents ranging from 10 to 660 mA (median [IQR]: preoperative \\u0026ndash; 30 [20\\u0026ndash;35]; postoperative \\u0026ndash; 35 [20\\u0026ndash;35]). Most preoperative and postoperative scans were performed without intravenous contrast.\\u003c/p\\u003e\\n\\u003ch3\\u003eTraining and inference data\\u003c/h3\\u003e\\n\\u003cp\\u003eThe dataset was divided into training and test sets ensuring that no patient had scans in both sets. A separate validation set was not required, as the self-configuring nnU-Net framework does not rely on manual hyperparameter tuning. The training pipelines for both the preoperative and postoperative segmentation models are illustrated in Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e. Conversion of DICOM files to NIfTI format de-identified the images by removing all protected health information from the image metadata.\\u003c/p\\u003e\\u003cp\\u003e\\u003c/p\\u003e\\u003cp\\u003eTo ensure generalizability, the preoperative lobe segmentation model was trained using two datasets: an internal institutional dataset and an external public dataset (OttawaChestCT dataset)\\u003csup\\u003e\\u003cspan citationid=\\\"CR21\\\" class=\\\"CitationRef\\\"\\u003e21\\u003c/span\\u003e\\u003c/sup\\u003e. The internal Stanford dataset included a random subset of 41 CT scans (34 for training and 7 for test) selected from a total of 236 clinical scans. The OttawaChestCT dataset contributed 81 scans (59 for training and 22 for test) out of the original 100; 19 scans were excluded due to incomplete lobe annotations (i.e., missing one or more of the five lobe masks). In total, the combined dataset comprised 122 annotated CT scans, with voxel-wise labels for six classes: background, right upper lobe (RUL), RML, right lower lobe (RLL), left upper lobe (LUL), and left lower lobe (LLL).\\u003c/p\\u003e\\u003cp\\u003eThe training pipeline for the postoperative lobe segmentation model utilized only the internal institutional dataset, consisting of randomly selected 82 CT scans (67 for training and 15 for test) from the full cohort of 236 clinical scans. Each scan was manually annotated with four segmentation classes: background, RML, RLL, and left lung.\\u003c/p\\u003e\\u003cp\\u003e\\u003cb\\u003eReference standard for segmentation masks\\u003c/b\\u003e\\u003c/p\\u003e\\u003cp\\u003eReference standard segmentation masks were generated using 3D Slicer (version 8.5.1; MIT, Massachusetts, USA) by D.N.K. for preoperative and postoperative CT scans in the internal Stanford dataset, while the original annotations provided by the dataset creators were used for the OttawaChestCT public dataset\\u003csup\\u003e\\u003cspan citationid=\\\"CR21\\\" class=\\\"CitationRef\\\"\\u003e21\\u003c/span\\u003e\\u003c/sup\\u003e.\\u003c/p\\u003e\\u003cdiv id=\\\"Sec8\\\" class=\\\"Section2\\\"\\u003e\\u003ch2\\u003eModel training\\u003c/h2\\u003e\\u003cp\\u003eThis study utilized the nnU-Net v2 model, a self-configuring deep learning framework for biomedical image segmentation. Its predecessor, nnU-Net v1, was built upon the U-Net architecture and designed to automatically adapt its network architecture, preprocessing steps, and training pipeline to the specific characteristics of a given dataset\\u0026mdash;eliminating the need for manual tuning of architectural or hyperparameter settings. The nnU-Net v2 retains the original self-adapting architecture and automated training configuration of the nnU-Net framework, with enhancements focused on improved training efficiency and broader device compatibility. Detailed implementation of the model is presented in the original paper\\u003csup\\u003e\\u003cspan citationid=\\\"CR22\\\" class=\\\"CitationRef\\\"\\u003e22\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e\\u003c/sup\\u003e.\\u003c/p\\u003e\\u003cp\\u003eModel training was performed using a single NVIDIA H100-SXM5 GPU. The 3D full-resolution nnU-Net v2 model was trained separately for preoperative and postoperative CT datasets, with each training run spanning 5,000 epochs (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e). Training time per epoch ranged from 22 to 30 seconds, resulting in a total training duration of approximately 42 hours for each model. Input CT images were resampled to a median voxel spacing of 1.25\\u0026times;0.70\\u0026times;0.70 mm and normalized using CT-specific intensity scaling. Due to the large volumetric size of CT scans in the dataset (median shape 258\\u0026times;512\\u0026times;512), training was performed with a patch size of 96\\u0026times;160\\u0026times;160 voxels and a batch size of 2. A Leaky ReLU activation function was used in all the hidden layers. The model was trained using the Adam optimizer with an initial learning rate of 0.001, following a polynomial (poly) learning rate decay schedule throughout training. Combined Dice and cross-entropy loss was used to guide optimization. Deep supervision was enabled. The checkpoint with the highest pseudo-Dice coefficient\\u0026mdash;i.e., the approximate Dice score computed on training patches rather than on the full CT volume\\u0026mdash;across all lung lobes on the training dataset was selected for further evaluation on both the test set and the full clinical dataset. Due to the absence of reference standard masks for the full clinical dataset, predicted lobe segmentations were visually inspected to confirm adequate anatomical accuracy.\\u003c/p\\u003e\\u003cp\\u003e\\u003c/p\\u003e\\u003c/div\\u003e\\n\\u003ch3\\u003ePreprocessing, data augmentation, and postprocessing\\u003c/h3\\u003e\\n\\u003cp\\u003eThese were handled using the default nnU-Net pipeline. Preprocessing included resampling CT scans to the median dataset voxel spacing, instance normalization using foreground intensity statistics, and cropping to nonzero regions to remove background. Data augmentation was performed on-the-fly and included random spatial transformation (rotation, scaling, elastic deformation) and intensity augmentations (gamma, noise, blur). Postprocessing involved removing all but the largest connected component for each predicted lobe label, applied selectively based on whether it improved the Dice score compared to baseline segmentation.\\u003c/p\\u003e\\u003cp\\u003eIn addition to the default nnU-Net postprocessing, customized automated postprocessing was applied to further refine the predicted masks. This included removal of anatomically discordant contralateral lobe predictions and isolated 3D islands disconnected from the largest lobar component. These customized postprocessing steps were not part of the training pipeline but were implemented post-inference and resulted in a marginal improvement in overall Dice scores.\\u003c/p\\u003e\\n\\u003ch3\\u003eVolumetric quantification of atelectasis\\u003c/h3\\u003e\\n\\u003cp\\u003ePreoperative and postoperative lung volumes for the full clinical dataset were quantified using automatically segmented masks derived from model inference on the respective CT scans. The preoperative CT segmented regions included RUL, RML, RLL, LUL, and LLL while the postoperative regions included RML, RLL, and left lung. Voxel volumes were calculated using the dimensions of each voxel. Lobar and total lung volumes were computed by multiplying the voxel volume (converted from mm\\u0026sup3; to mL) with the number of voxels in each label. A subset of the calculated volumes was cross validated against measurements obtained using the standard volume computation tool in 3D Slicer (version 8.5.1; MIT, Massachusetts, USA) to ensure accuracy. RML and RLL volumes were also normalized with right lung and total lung volumes to minimize interscan variability.\\u003c/p\\u003e\\u003cdiv id=\\\"Sec11\\\" class=\\\"Section2\\\"\\u003e\\u003ch2\\u003eStatistical analysis\\u003c/h2\\u003e\\u003cp\\u003eDescriptive statistics were used to summarize baseline demographic and clinical data. The normality of continuous data was evaluated using the Shapiro-Wilk test. Normally distributed continuous variables were reported as mean\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;standard deviation (SD) and compared using unpaired t-tests, while non-normally distributed variables were reported as median (interquartile range, IQR) and compared using the Mann-Whitney U test. Categorical variables were presented as frequencies and percentages and compared using Pearson\\u0026rsquo;s chi-square test (χ\\u0026sup2;) or Fisher\\u0026rsquo;s exact test, as appropriate. The Jonckheere\\u0026ndash;Terpstra test was used to assess for ordered trends in all volume metrics across atelectasis severity grades. A two-tailed α\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.05 was considered statistically significant. Bonferroni correction was applied for multiple testing. All analyses were conducted using SPSS software (version 29.0, IBM Corp., Armonk, NY, USA), R (version 4.4.2, R Foundation for Statistical Computing, Vienna, Austria), and Python (version 3.13, Python Software Foundation, Wilmington, DE, USA).\\u003c/p\\u003e\\u003c/div\\u003e\"},{\"header\":\"Results\",\"content\":\"\\u003cdiv id=\\\"Sec13\\\" class=\\\"Section2\\\"\\u003e\\u003ch2\\u003eBaseline Cohort Characteristics\\u003c/h2\\u003e\\u003cp\\u003eOut of the 438 patients who underwent right upper lobectomy, 236 patients were included in the study after excluding those without available preoperative or postoperative scans within the defined time windows. Baseline demographic and clinical characteristics are summarized in Table\\u0026nbsp;\\u003cspan refid=\\\"Tab1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e. The median age was 69.2 years (IQR: 61.0\\u0026ndash;74.6) and 62.3% of patients were female. Common comorbidities included diabetes mellitus (17.8%), chronic obstructive pulmonary disease (16.5%), cerebrovascular disease (11.4%), and chronic kidney disease (11.0%).\\u003c/p\\u003e\\u003cp\\u003e\\u003cdiv class=\\\"gridtable\\\"\\u003e\\u003ctable float=\\\"Yes\\\" id=\\\"Tab1\\\" border=\\\"1\\\"\\u003e\\u003ccaption language=\\\"En\\\"\\u003e\\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 1\\u003c/div\\u003e\\u003cdiv class=\\\"CaptionContent\\\"\\u003e\\u003cp\\u003eBaseline demographic and clinical characteristics of all patients who underwent right upper lobectomy and those who were included in the final study cohort.\\u003c/p\\u003e\\u003c/div\\u003e\\u003c/caption\\u003e\\u003ccolgroup cols=\\\"4\\\"\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c1\\\" colnum=\\\"1\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c2\\\" colnum=\\\"2\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c3\\\" colnum=\\\"3\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c4\\\" colnum=\\\"4\\\"\\u003e\\u003c/div\\u003e\\u003cthead\\u003e\\u003ctr\\u003e\\u003cth align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c2\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003eCharacteristics\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003eRight upper lobectomy cohort\\u003c/p\\u003e\\u003cp\\u003e(n\\u0026thinsp;=\\u0026thinsp;438)\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003eStudy cohort\\u003c/p\\u003e\\u003cp\\u003e(n\\u0026thinsp;=\\u0026thinsp;236)\\u003c/p\\u003e\\u003c/th\\u003e\\u003c/tr\\u003e\\u003c/thead\\u003e\\u003ctbody\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c2\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003eAge (years) (Median, IQR)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e69.7(62.7\\u0026ndash;75.0)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e69.2(61.0-74.6)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\" morerows=\\\"1\\\" rowspan=\\\"2\\\"\\u003e\\u003cp\\u003eSex\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eMale\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e179(40.9%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e89(37.9%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eFemale\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e259(59.1%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e147(62.3%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\" morerows=\\\"4\\\" rowspan=\\\"5\\\"\\u003e\\u003cp\\u003eRace\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eWhite\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e282(64.4%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e144(61.0%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eBlack\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e7(1.6%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e5(2.1%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eAsian\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e98(22.4%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e64(27.1%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eOther\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e43(9.8%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e23(9.7%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eUnknown\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e8(1.8%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0(0%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\" morerows=\\\"2\\\" rowspan=\\\"3\\\"\\u003e\\u003cp\\u003eApproach\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eOpen\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e117(26.7%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e50(21.2%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eVATS\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e248(56.6%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e134(56.8%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eRobot-assisted\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e73(16.7%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e52(22.0%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c2\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003eFEV\\u003csub\\u003e1\\u003c/sub\\u003e (%) (Median, IQR)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e95(79.2\\u0026ndash;108)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e97(80\\u0026ndash;109)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c2\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003eDiabetes mellitus\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e64(14.6%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e42(17.8%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c2\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003eChronic obstructive pulmonary disease\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e70(16.0%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e39(16.5%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c2\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003eCongestive heart failure\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e25(5.7%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e15(6.4%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c2\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003eHistory of cerebrovascular accident\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e45(10.3%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e27(11.4%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c2\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003eChronic kidney disease\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e40(9.1%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e26(11.0%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c2\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003eCharlson-comorbidity score (Median, IQR)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e6(4\\u0026ndash;7)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e6(5\\u0026ndash;8)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\" morerows=\\\"6\\\" rowspan=\\\"7\\\"\\u003e\\u003cp\\u003eOperative Diagnosis\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eAdenocarcinoma\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e329(75.1%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e191(80.9%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eSCC\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e43(9.8%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e14(5.9%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eNEC\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e16(3.7%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e6(2.5%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eMixed\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e10(2.3%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e3(1.3%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eMetastases\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e17(3.9%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e11(4.7%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eGranulomatous inflammation\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e7(1.6%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e3(1.3%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eOther benign lesions\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e16(3.7%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e8(3.4%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c2\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003eMedian time from preop scan to surgery (days) (Median, IQR)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c4\\\" namest=\\\"c3\\\"\\u003e\\u003cp\\u003e44.5(24.0\\u0026ndash;74.0)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c2\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003eMedian time from surgery to postop scan (days) (Median, IQR)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c4\\\" namest=\\\"c3\\\"\\u003e\\u003cp\\u003e196(173.2-207.7)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\" morerows=\\\"4\\\" rowspan=\\\"5\\\"\\u003e\\u003cp\\u003eAtelectasis on CT scan at 6 months\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e260(78.3%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e186 (78.8%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e1\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e48(14.5%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e32 (13.6%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e2\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e13(3.9%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e10 (4.2%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e3\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e6(1.8%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e6 (2.5%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e4\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e5(1.5%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e2 (0.8%)\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003c/tbody\\u003e\\u003c/colgroup\\u003e\\u003c/table\\u003e\\u003c/div\\u003e\\u003c/p\\u003e\\u003c/div\\u003e\\u003cdiv id=\\\"Sec14\\\" class=\\\"Section2\\\"\\u003e\\u003ch2\\u003eRadiological data characteristics\\u003c/h2\\u003e\\u003cp\\u003eThe median time between preoperative CT scans and the surgery was 44.5 (24, 74) days and the median time between the surgery and postoperative CT scans was 196 (173.2, 207.7) days. The median slice thickness was 1.25 (1.0, 1.25) mm for preoperative and postoperative scans. RML atelectasis at 6 months was absent in 186 (78.8%) of patients, while grade 1, 2, 3, and 4 RML atelectasis were observed in 32 (13.6%), 10 (4.2%), 6 (2.5%), and 2 (0.8%) patients.\\u003c/p\\u003e\\u003c/div\\u003e\\u003cdiv id=\\\"Sec15\\\" class=\\\"Section2\\\"\\u003e\\u003ch2\\u003ePreoperative model performance\\u003c/h2\\u003e\\u003cp\\u003eOn the training dataset (n\\u0026thinsp;=\\u0026thinsp;93), Dice scores ranged from 0.97\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01 for the RML to 0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.00 for the RLL, with an overall mean Dice of 0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01. The internal test dataset (n\\u0026thinsp;=\\u0026thinsp;29) showed similarly strong performance, achieving an overall mean Dice score of 0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02 after postprocessing (Table\\u0026nbsp;\\u003cspan refid=\\\"Tab2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e). The RML consistently yielded the lowest Dice scores across training and test datasets, likely reflecting the higher likelihood of incomplete right horizontal fissure vs. other fissures and smaller volume of this lobe.\\u003c/p\\u003e\\u003cp\\u003e\\u003cdiv class=\\\"gridtable\\\"\\u003e\\u003ctable float=\\\"Yes\\\" id=\\\"Tab2\\\" border=\\\"1\\\"\\u003e\\u003ccaption language=\\\"En\\\"\\u003e\\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 2\\u003c/div\\u003e\\u003cdiv class=\\\"CaptionContent\\\"\\u003e\\u003cp\\u003ePreoperative and postoperative lung segmentation model performance on CT scans, reported as Dice scores for each lung lobe and overall mean Dice score.\\u003c/p\\u003e\\u003c/div\\u003e\\u003c/caption\\u003e\\u003ccolgroup cols=\\\"6\\\"\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c1\\\" colnum=\\\"1\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"char\\\" char=\\\"\\u0026plusmn;\\\" class=\\\"colspec\\\" colname=\\\"c2\\\" colnum=\\\"2\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c3\\\" colnum=\\\"3\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c4\\\" colnum=\\\"4\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c5\\\" colnum=\\\"5\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c6\\\" colnum=\\\"6\\\"\\u003e\\u003c/div\\u003e\\u003cthead\\u003e\\u003ctr\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c1\\\" morerows=\\\"1\\\" rowspan=\\\"2\\\"\\u003e\\u003cp\\u003ePreoperative\\u003c/p\\u003e\\u003cp\\u003eLung lobes\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c2\\\" morerows=\\\"1\\\" rowspan=\\\"2\\\"\\u003e\\u003cp\\u003eTraining dataset\\u003c/p\\u003e\\u003cp\\u003e(93)\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c4\\\" namest=\\\"c3\\\"\\u003e\\u003cp\\u003eInternal test\\u003c/p\\u003e\\u003cp\\u003edataset (29)\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c6\\\" namest=\\\"c5\\\"\\u003e\\u003cp\\u003eExternal validation dataset (55)\\u003c/p\\u003e\\u003c/th\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003eModel inference\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003eAfter post-processing\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003eModel inference\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003eAfter post-processing\\u003c/p\\u003e\\u003c/th\\u003e\\u003c/tr\\u003e\\u003c/thead\\u003e\\u003ctbody\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eRight upper lobe\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\"\\u0026plusmn;\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.97\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.97\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.94\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.07\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.95\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.06\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eRight middle lobe\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\"\\u0026plusmn;\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.97\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.96\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.96\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.86\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.22\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.86\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.22\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eRight lower lobe\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\"\\u0026plusmn;\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.04\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.93\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.12\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.94\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.12\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eLeft upper lobe\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\"\\u0026plusmn;\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.00\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.93\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.13\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.95\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;.13\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eLeft lower lobe\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\"\\u0026plusmn;\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.89\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.18\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.91\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.18\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eOverall Mean\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\"\\u0026plusmn;\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.97\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.91\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.16\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e0.92\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.15\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\" morerows=\\\"1\\\" rowspan=\\\"2\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003ePostoperative\\u003c/b\\u003e\\u003c/p\\u003e\\u003cp\\u003e\\u003cb\\u003eLung lobes\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\" morerows=\\\"1\\\" rowspan=\\\"2\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eTraining dataset\\u003c/b\\u003e\\u003c/p\\u003e\\u003cp\\u003e\\u003cb\\u003e(67)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c4\\\" namest=\\\"c3\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eInternal test\\u003c/b\\u003e\\u003c/p\\u003e\\u003cp\\u003e\\u003cb\\u003edataset (15)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c6\\\" namest=\\\"c5\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eExternal validation dataset\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eModel inference\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eAfter post-processing\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e-\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e-\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eRight middle lobe\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\"\\u0026plusmn;\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.96\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.1\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e-\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e-\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eRight lower lobe\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\"\\u0026plusmn;\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.00\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.00\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e-\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e-\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eLeft lung\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\"\\u0026plusmn;\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.00\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.00\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e-\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e-\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eOverall Mean\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\"\\u0026plusmn;\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.04\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.00\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.00\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e-\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e-\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003c/tbody\\u003e\\u003c/colgroup\\u003e\\u003c/table\\u003e\\u003c/div\\u003e\\u003c/p\\u003e\\u003cp\\u003eOn the external test dataset of LOLA11 challenge (n\\u0026thinsp;=\\u0026thinsp;55), baseline model performance showed moderate variability across lobes, with Dice scores ranging from 0.86\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.22 for the RML to 0.94\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.07 for the RUL. Postprocessing yielded minor improvements in the mean Dice score from 0.91\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.16 to 0.923\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.15. This score currently is the joint state-of-the-art in the most comprehensive and longest running lung lobe segmentation challenge (0.928\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.15) \\u003csup\\u003e\\u003cspan citationid=\\\"CR24\\\" class=\\\"CitationRef\\\"\\u003e24\\u003c/span\\u003e\\u003c/sup\\u003e.\\u003c/p\\u003e\\u003c/div\\u003e\\u003cdiv id=\\\"Sec16\\\" class=\\\"Section2\\\"\\u003e\\u003ch2\\u003ePostoperative model performance\\u003c/h2\\u003e\\u003cp\\u003eOn the training dataset (n\\u0026thinsp;=\\u0026thinsp;67), Dice scores ranged from 0.96\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.1 for the RML to 0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02 for the left lung, with an overall mean Dice score of 0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.04 (Table\\u0026nbsp;\\u003cspan refid=\\\"Tab2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e). Internal testing (n\\u0026thinsp;=\\u0026thinsp;15) yielded excellent performance, with an overall Dice score of 0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.00 after postprocessing. Among the individual lobes, the RML showed slightly lower Dice scores.\\u003c/p\\u003e\\u003c/div\\u003e\\u003cdiv id=\\\"Sec17\\\" class=\\\"Section2\\\"\\u003e\\u003ch2\\u003eAutomated inference, volumetric analysis, and atelectasis grading in clinical CT dataset\\u003c/h2\\u003e\\u003cp\\u003eWe applied our preoperative and post-operative model inference to the entire clinical CT dataset (n\\u0026thinsp;=\\u0026thinsp;236). Eight postoperative scans (3.4%) were excluded from analysis owing to pleural pathologies, including effusion and pneumothorax. Owing to the lack of a reference standard in the postoperative dataset, all output masks were visually inspected, and manual corrections were required in 16 scans (6.8%) to ensure accurate lobar volumetry for atelectasis assessment. The median Dice score between the edited and original masks (n\\u0026thinsp;=\\u0026thinsp;16) was 0.85 (IQR, 0.74\\u0026ndash;0.94).\\u003c/p\\u003e\\u003cp\\u003eWe then extracted lobar volumes and assessed longitudinal lobar volume change using various metrics (Table\\u0026nbsp;\\u003cspan refid=\\\"Tab3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003e). Median (IQR) ΔRML volume change demonstrated a progressive decline with increasing RML atelectasis grades from \\u0026minus;\\u0026thinsp;4.6 mL (-78.5, 59.0) (no atelectasis) to \\u0026minus;\\u0026thinsp;317.8 mL (-440.7, -194.8) (grade 4 atelectasis) (p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.001). This pattern persisted across normalized metrics, with ΔRML/RL and ΔRML/TL volume changes showing increasingly negative values with increasing grades of atelectasis (p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.001). Importantly, RLL volume changes showed an opposite trend, highlighting the expected compensatory hyperinflation of the RLL with increasing degrees of RML atelectasis. ΔRLL/RL volumes increased with worsening RML atelectasis (e.g., median ΔRLL/RL \\u0026minus;\\u0026thinsp;31.2% to 50.5%, p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.001) (Table\\u0026nbsp;\\u003cspan refid=\\\"Tab3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003e). However, the ΔRLL/TL did not show a similar trend (p\\u0026thinsp;=\\u0026thinsp;0.79). While the overall trend of RML volume loss and compensatory ΔRLL/RL volume increase was significant, pairwise comparisons reached statistical significance only between grades 0\\u0026ndash;1 and occasionally in grades 1\\u0026ndash;2, but not between higher grades (2\\u0026ndash;3 or 3\\u0026ndash;4), likely due to low sample sizes in those categories (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003eA-E).\\u003c/p\\u003e\\u003cp\\u003e\\u003cdiv class=\\\"gridtable\\\"\\u003e\\u003ctable float=\\\"Yes\\\" id=\\\"Tab3\\\" border=\\\"1\\\"\\u003e\\u003ccaption language=\\\"En\\\"\\u003e\\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 3\\u003c/div\\u003e\\u003cdiv class=\\\"CaptionContent\\\"\\u003e\\u003cp\\u003eQuantitative assessment of lobar volume changes between preoperative and 6-month postoperative CT scans, stratified by radiographic atelectasis grade of 0\\u0026ndash;4. Volume changes are presented as absolute (mL) and percentage changes (%) relative to right lung (RL) and total lung (TL) volumes (Median, IQR).\\u003c/p\\u003e\\u003c/div\\u003e\\u003c/caption\\u003e\\u003ccolgroup cols=\\\"7\\\"\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c1\\\" colnum=\\\"1\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c2\\\" colnum=\\\"2\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c3\\\" colnum=\\\"3\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c4\\\" colnum=\\\"4\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c5\\\" colnum=\\\"5\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c6\\\" colnum=\\\"6\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"char\\\" char=\\\".\\\" class=\\\"colspec\\\" colname=\\\"c7\\\" colnum=\\\"7\\\"\\u003e\\u003c/div\\u003e\\u003cthead\\u003e\\u003ctr\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003eMetrics\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eNone\\u003c/p\\u003e\\u003cp\\u003e(0)\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003eMinimal\\u003c/p\\u003e\\u003cp\\u003e(1)\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003eSub-segmental (2)\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003eSegmental (3)\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003eLobar\\u003c/p\\u003e\\u003cp\\u003e(4)\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003ep-value\\u003c/p\\u003e\\u003cp\\u003e(for trend)\\u003c/p\\u003e\\u003c/th\\u003e\\u003c/tr\\u003e\\u003c/thead\\u003e\\u003ctbody\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eMedian RML volume change (mL)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e-4.6 (-78.5, 59.0)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e-74.3 (-146.5, -40.5)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e-189.9 (-296.2, -96.9)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e-244.9 (-289.6, -199.2)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e-317.8 (-440.7, -194.8)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e\\u0026lt;\\u0026thinsp;0.001\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eMedian RML/RL volume change (%)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e3.7 (1.1, 6.5)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.7 (-2.3, 3.2)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e-4.5 (-8.6, -1.1)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e-9.2 (-12.1, -5.1)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e-9.5 (-12.1, -7.0)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e\\u0026lt;\\u0026thinsp;0.001\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eMedian RML/TL volume change (%)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.6 (-0.9, 1.8)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e-1.7 (-2.7, -0.2)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e-3.1 (-5.2, -1.5)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e-5.2 (-6.7, -3.5)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e-5.1 (-6.4, -3.7)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e\\u0026lt;\\u0026thinsp;0.001\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eMedian RLL/RL volume change (%)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e31.2 (27.5, 36.3)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e34.0 (29.4, 37.1)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e38.2 (36.0, 42.5)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e42.7 (35.6, 47.1)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e50.5 (49.2, 51.9)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e\\u0026lt;\\u0026thinsp;0.001\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eMedian RLL/TL volume change (%)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e11.6 (9.5, 14.0)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e11.1 (5.8, 13.4)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e12.4 (9.1, 15.3)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e13.0 (6.0, 19.5)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e17.5 (17.1, 18.0)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.79\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003c/tbody\\u003e\\u003c/colgroup\\u003e\\u003ctfoot\\u003e\\u003ctr\\u003e\\u003ctd colspan=\\\"7\\\"\\u003eRML, right middle lobe; RLL, right lower lobe; RL, right lung; TL, total lung.\\u003c/td\\u003e\\u003c/tr\\u003e\\u003c/tfoot\\u003e\\u003c/table\\u003e\\u003c/div\\u003e\\u003c/p\\u003e\\u003cp\\u003e\\u003c/p\\u003e\\u003cp\\u003eWhile individual comparisons between consecutive atelectasis grades did not reach statistical significance, comparisons after pooling the grades of atelectasis as minimal-subsegmental (grades 1\\u0026ndash;2) and segmental-lobar collapse (grades 3\\u0026ndash;4) showed a statistically significant difference (Table\\u0026nbsp;\\u003cspan refid=\\\"Tab4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e). Notably, changes in ΔRML, ΔRML/RL, and ΔRML/TL between grades 0 vs 1\\u0026ndash;2 and 1\\u0026ndash;2 vs 3\\u0026ndash;4 were highly significant (p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.001). And while ΔRLL/RL across atelectasis grades also reached significance, ΔRLL/TL did not (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003eF-J).\\u003c/p\\u003e\\u003cp\\u003e\\u003cdiv class=\\\"gridtable\\\"\\u003e\\u003ctable float=\\\"Yes\\\" id=\\\"Tab4\\\" border=\\\"1\\\"\\u003e\\u003ccaption language=\\\"En\\\"\\u003e\\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 4\\u003c/div\\u003e\\u003cdiv class=\\\"CaptionContent\\\"\\u003e\\u003cp\\u003eQuantitative assessment of lobar volume changes between preoperative and 6-month postoperative CT scans, stratified by pooled atelectasis grades (grade 0, 1\\u0026ndash;2, and 3\\u0026ndash;4) Volume changes are presented as absolute (mL) and percentage changes (%) relative to right lung (RL) and total lung (TL) volumes (p-values for pairwise comparisons in Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003eF-J) (Median, IQR).\\u003c/p\\u003e\\u003c/div\\u003e\\u003c/caption\\u003e\\u003ccolgroup cols=\\\"5\\\"\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c1\\\" colnum=\\\"1\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c2\\\" colnum=\\\"2\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c3\\\" colnum=\\\"3\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c4\\\" colnum=\\\"4\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"char\\\" char=\\\".\\\" class=\\\"colspec\\\" colname=\\\"c5\\\" colnum=\\\"5\\\"\\u003e\\u003c/div\\u003e\\u003cthead\\u003e\\u003ctr\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003eMetrics\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eNone (0)\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003eMinimal (1)\\u003c/p\\u003e\\u003cp\\u003eAND\\u003c/p\\u003e\\u003cp\\u003eSub-segmental (2)\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003eSegmental (3)\\u003c/p\\u003e\\u003cp\\u003eAND\\u003c/p\\u003e\\u003cp\\u003eLobar (4)\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003ep-value\\u003c/p\\u003e\\u003cp\\u003e(for trend)\\u003c/p\\u003e\\u003c/th\\u003e\\u003c/tr\\u003e\\u003c/thead\\u003e\\u003ctbody\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eMedian RML volume change (mL)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e-4.4 (-78.6, 59.0)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e-93.7 (-168.6, -43.7)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e-244.9 (-318.0, -190.9)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e\\u0026lt;\\u0026thinsp;0.001\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eMedian RML/RL volume change (%)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e3.7 (1.1, 6.5)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e-0.5 (-3.5, 2.6)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e-9.2 (-12.9, -4.5)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e\\u0026lt;\\u0026thinsp;0.001\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eMedian RML/TL volume change (%)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.6 (-0.9, 1.8)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e-1.7 (-3.1, -0.4)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e-5.2 (-7.0, -3.3)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e\\u0026lt;\\u0026thinsp;0.001\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eMedian RLL/RL volume change (%)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e31.2 (27.5, 36.3)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e34.3 (29.6, 38.5)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e45.4 (40.2, 48.8)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e\\u0026lt;\\u0026thinsp;0.001\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eMedian RLL/TL volume change (%)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e11.6 (9.5, 14.0)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e11.1 (6.1, 13.6)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e17.2 (7.5, 18.9)\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.72\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003c/tbody\\u003e\\u003c/colgroup\\u003e\\u003ctfoot\\u003e\\u003ctr\\u003e\\u003ctd colspan=\\\"5\\\"\\u003eRML, right middle lobe; RLL, right lower lobe; RL, right lung; TL, total lung.\\u003c/td\\u003e\\u003c/tr\\u003e\\u003c/tfoot\\u003e\\u003c/table\\u003e\\u003c/div\\u003e\\u003c/p\\u003e\\u003cp\\u003e\\u003cb\\u003eSensitivity analysis of segmentation accuracy and volumetric assessments after excluding RML wedge resections and replacing the original unedited outputs\\u003c/b\\u003e\\u003c/p\\u003e\\u003cp\\u003eExcluding cases with additional RML wedge resections beyond the index RULobectomy procedure, which could confound the interpretation of lobar volume loss, resulted in similar directional trends in volume metrics, reinforcing the association between automated volume metrics and physician-assigned atelectasis severity grades (Supplementary Fig.\\u0026nbsp;2).\\u003c/p\\u003e\\u003cp\\u003eMoreover, analysis using the unedited segmentation outputs for the postoperative dataset yielded similar statistical differences in most volume metrics across both atelectasis and pooled atelectasis grades (Supplementary Fig.\\u0026nbsp;3).\\u003c/p\\u003e\\u003c/div\\u003e\\u003cdiv id=\\\"Sec18\\\" class=\\\"Section2\\\"\\u003e\\u003ch2\\u003eQualitative analysis\\u003c/h2\\u003e\\u003cp\\u003eVarious grades of RML atelectasis are illustrated in Figs.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e and \\u003cspan refid=\\\"Fig5\\\" class=\\\"InternalRef\\\"\\u003e5\\u003c/span\\u003e. Qualitative error analysis on segmentation errors was performed on CT scans where the Dice score was \\u0026lt;\\u0026thinsp;0.97. The most frequent errors were observed along the fissure between the RUL-RML and at the hilar junction of RML\\u0026ndash;RLL. Additional inaccuracies included over-segmentation of solid regions within dense consolidation areas, and under-segmentation of bullae (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig6\\\" class=\\\"InternalRef\\\"\\u003e6\\u003c/span\\u003e).\\u003c/p\\u003e\\u003cp\\u003eA detailed qualitative error analysis was performed on the external validation dataset, which is the most diverse publicly available dataset of chest CT scans for lung lobe segmentation (Fig.\\u0026nbsp;9). Representative errors over complex anatomical variations and challenging pathologies, such as scoliosis, chest wall deformities, severe COPD, large pleural effusions with near-complete lung collapse, and other abnormalities are illustrated in Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig7\\\" class=\\\"InternalRef\\\"\\u003e7\\u003c/span\\u003e.\\u003c/p\\u003e\\u003c/div\\u003e\"},{\"header\":\"Discussion\",\"content\":\"\\u003cp\\u003eIn this study, we developed and validated an automated, deep learning-based volumetric pipeline to quantify and grade atelectasis\\u0026ndash;one of the most common clinical problems in pulmonary medicine and surgery. We selected postoperative atelectasis of the middle lobe following right upper lobectomy as a representative model to demonstrate that atelectasis can be graded using lobar volumetry. Importantly, such an automated system to reliably quantify atelectasis would also likely be applicable to the broader assessment of atelectasis across diverse clinical scenarios.\\u003c/p\\u003e\\u003cp\\u003eAlthough deep learning\\u0026ndash;based models have demonstrated high accuracy in delineating lobar anatomy with Dice scores of 92\\u0026ndash;96%\\u003csup\\u003e25\\u0026ndash;28\\u003c/sup\\u003e (Table\\u0026nbsp;\\u003cspan refid=\\\"Tab5\\\" class=\\\"InternalRef\\\"\\u003e5\\u003c/span\\u003e), they have been trained on preoperative anatomically normal lungs, with limited validation in postoperative settings where anatomical distortion is common. To date, only one study assessed lobar segmentation before and after various lobectomies, but without examining associations with clinical outcomes\\u003csup\\u003e\\u003cspan citationid=\\\"CR29\\\" class=\\\"CitationRef\\\"\\u003e29\\u003c/span\\u003e\\u003c/sup\\u003e. Our models achieved high accuracy in delineating lobar anatomy, even in the anatomically distorted postoperative setting, and further demonstrated that volumetric reduction of the RML with compensatory hyperinflation of RLL correlated with increasing clinically determined grades of atelectasis. These findings remained consistent in sensitivity analyses. Together, the results support feasibility of automated lobar volumetry as an objective surrogate for grading atelectasis severity, with the caveat that some cases were excluded and some required manual adjustment.\\u003c/p\\u003e\\u003cp\\u003e\\u003cdiv class=\\\"gridtable\\\"\\u003e\\u003ctable float=\\\"Yes\\\" id=\\\"Tab5\\\" border=\\\"1\\\"\\u003e\\u003ccaption language=\\\"En\\\"\\u003e\\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 5\\u003c/div\\u003e\\u003cdiv class=\\\"CaptionContent\\\"\\u003e\\u003cp\\u003ePrior state-of-the-art lung lobe segmentation performance reported in the literature and on the publicly available LOLA11 challenge dataset comprising 55 chest CT scans.\\u003c/p\\u003e\\u003c/div\\u003e\\u003c/caption\\u003e\\u003ccolgroup cols=\\\"7\\\"\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c1\\\" colnum=\\\"1\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c2\\\" colnum=\\\"2\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c3\\\" colnum=\\\"3\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c4\\\" colnum=\\\"4\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c5\\\" colnum=\\\"5\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c6\\\" colnum=\\\"6\\\"\\u003e\\u003c/div\\u003e\\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c7\\\" colnum=\\\"7\\\"\\u003e\\u003c/div\\u003e\\u003cthead\\u003e\\u003ctr\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003eAuthor (Year) [citation]\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003eRight upper lobe\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003eRight middle lobe\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003eRight lower lobe\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003eLeft upper lobe\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003eLeft lower lobe\\u003c/p\\u003e\\u003c/th\\u003e\\u003cth align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003eOverall\\u003c/p\\u003e\\u003c/th\\u003e\\u003c/tr\\u003e\\u003c/thead\\u003e\\u003ctbody\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eDoel et al. (2012) \\u0026ndash; Traditional CV methods\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR34\\\" class=\\\"CitationRef\\\"\\u003e34\\u003c/span\\u003e\\u003c/sup\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.86\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.15\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.55\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.40\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.77\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.34\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.88\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.21\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.86\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.24\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.79\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.31\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e3D U-Net (2016\\u0026ndash;2020)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR36\\\" citationid=\\\"CR35\\\" class=\\\"CitationRef\\\"\\u003e35\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR37\\\" class=\\\"CitationRef\\\"\\u003e37\\u003c/span\\u003e\\u003c/sup\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.93\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.07\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.84\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.12\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.93\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.04\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.92\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.04\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.93\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.04\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.91\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.08\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eFerreira et al. FRV-Net (2018)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR38\\\" class=\\\"CitationRef\\\"\\u003e38\\u003c/span\\u003e\\u003c/sup\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.94\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.07\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.87\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.12\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.94\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.05\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.95\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.95\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.04\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.92\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eGerard et al. Series of 3D CNN (2019)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR39\\\" class=\\\"CitationRef\\\"\\u003e39\\u003c/span\\u003e\\u003c/sup\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.99\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.98\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.99\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.99\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.99\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.99\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eXie et al. RU-Net (2020)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR27\\\" class=\\\"CitationRef\\\"\\u003e27\\u003c/span\\u003e\\u003c/sup\\u003e \\u003cb\\u003e\\u0026ndash; (IoU)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.95\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.96\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.96\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.96\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.96\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.95\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eZheng et al. (2021) - Dual attention network\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR40\\\" class=\\\"CitationRef\\\"\\u003e40\\u003c/span\\u003e\\u003c/sup\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.93\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.90\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.96\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.94\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.94\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.01\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.93\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eZhang et al. (2021)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR41\\\" class=\\\"CitationRef\\\"\\u003e41\\u003c/span\\u003e\\u003c/sup\\u003e\\u003c/p\\u003e\\u003cp\\u003e\\u003cb\\u003eDenseVNet\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.96\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.92\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.96\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.94\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.93\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.94\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003ePeng et al. (2022)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR28\\\" class=\\\"CitationRef\\\"\\u003e28\\u003c/span\\u003e\\u003c/sup\\u003e \\u003cb\\u003e- Multi-feature fusion and Ensemble.\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.97\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.88\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.96\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.98\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.97\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.95\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eBao et al. (2023)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR42\\\" class=\\\"CitationRef\\\"\\u003e42\\u003c/span\\u003e\\u003c/sup\\u003e \\u003cb\\u003eEdge enhancement cascaded network\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.98\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.96\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.98\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.98\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.97\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.97\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eNomura et al (2025)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR43\\\" class=\\\"CitationRef\\\"\\u003e43\\u003c/span\\u003e\\u003c/sup\\u003e \\u003cb\\u003eTriSwinUNETR\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.93\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.85\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.95\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.97\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.96\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.93\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colspan=\\\"7\\\" nameend=\\\"c7\\\" namest=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eLOLA11 lung lobe segmentation challenge\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR24\\\" class=\\\"CitationRef\\\"\\u003e24\\u003c/span\\u003e\\u003c/sup\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eVan Rikxkoort et al. (2010)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR44\\\" class=\\\"CitationRef\\\"\\u003e44\\u003c/span\\u003e\\u003c/sup\\u003e \\u003cb\\u003eTraditional CV\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u0026nbsp;\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u0026nbsp;\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u0026nbsp;\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u0026nbsp;\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u0026nbsp;\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.85\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eLassen et al. (2013)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR45\\\" class=\\\"CitationRef\\\"\\u003e45\\u003c/span\\u003e\\u003c/sup\\u003e \\u003cb\\u003eTraditional CV\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u0026nbsp;\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u0026nbsp;\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u0026nbsp;\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u0026nbsp;\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u0026nbsp;\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.88\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eBragman et al. (2017)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR25\\\" class=\\\"CitationRef\\\"\\u003e25\\u003c/span\\u003e\\u003c/sup\\u003e \\u003cb\\u003eProbabilistic model\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.91\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.20\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.88\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.24\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.93\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.07\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.80\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.23\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.91\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.19\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.88\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eImran et al. (2018)\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR26\\\" class=\\\"CitationRef\\\"\\u003e26\\u003c/span\\u003e\\u003c/sup\\u003e \\u003cb\\u003ePDV-Net\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.93\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.07\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.86\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.12\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.95\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.93\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.94\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.03\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e0.92\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.07\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eCurrent SOTA \\u0026ndash; nnU-Net (2025)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e0.95\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.08\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e0.86\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.22\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.96\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.06\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e0.95\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.13\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.92\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.19\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e0.928\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.15\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003ctr\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003eOurs (2025)\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e0.95\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.06\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e0.86\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.22\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e\\u003cp\\u003e0.94\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.13\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e0.95\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.13\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e\\u003cp\\u003e0.91\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.17\\u003c/p\\u003e\\u003c/td\\u003e\\u003ctd align=\\\"left\\\" colname=\\\"c7\\\"\\u003e\\u003cp\\u003e\\u003cb\\u003e0.923\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.16\\u003c/b\\u003e\\u003c/p\\u003e\\u003c/td\\u003e\\u003c/tr\\u003e\\u003c/tbody\\u003e\\u003c/colgroup\\u003e\\u003c/table\\u003e\\u003c/div\\u003e\\u003c/p\\u003e\\u003cp\\u003eDespite claims of superior performance of novel architectures\\u0026mdash;transformer-based and Mamba-based\\u0026ndash;over CNNs for 3D medical segmentation, the CNN-based nnU-Net demonstrated the best Dice scores in a recent comprehensive benchmarking of six 3D medical datasets\\u003csup\\u003e\\u003cspan citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e\\u003c/sup\\u003e. In addition to its robust and generalizable performance with relatively small datasets and its self-configuring architecture, the added advantages of lower VRAM consumption and shorter training times vs. other architectures influenced our choice of nnU-Net for model training. Table\\u0026nbsp;\\u003cspan refid=\\\"Tab5\\\" class=\\\"InternalRef\\\"\\u003e5\\u003c/span\\u003e lists other preoperative lung lobe segmentation models reporting Dice scores up to 0.99 but which lack validation on external datasets like LOLA11\\u003csup\\u003e24\\u003c/sup\\u003e.\\u003c/p\\u003e\\u003cp\\u003eMoreover, our study demonstrates that deep learning models can robustly segment even complex, anatomically distorted postoperative lung anatomy following lobectomy. We have also demonstrated the ability of the models to generalize across vendors/scanners, acquisition protocols, and reconstruction kernels. While the postoperative model was generally robust, ~\\u0026thinsp;6% of scans required manual editing to ensure accurate volumetry\\u0026mdash;typically in cases with more severe anatomical distortion. This highlights the potential for improved performance through training on larger, diverse datasets.\\u003c/p\\u003e\\u003cp\\u003eTo our knowledge, this is the first study to apply DL\\u0026ndash;based analysis for grading postoperative atelectasis. While prior work has attempted binary classification of atelectasis presence\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR31\\\" citationid=\\\"CR30\\\" class=\\\"CitationRef\\\"\\u003e30\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR32\\\" class=\\\"CitationRef\\\"\\u003e32\\u003c/span\\u003e\\u003c/sup\\u003e, such approaches are overly simplistic and fail to reflect the true spectrum of disease severity. Moreover, manual grading systems are prone to mild-moderate inter- and intra-rater variability, limiting their reliability and scalability. Our approach addresses these limitations by not only detecting presence of atelectasis as a \\u0026ldquo;present/absent\\u0026rdquo; decision, but also assessing its severity in a more objective, continuous, and reproducible manner by quantifying lobar volume loss. While our study focused only on RML atelectasis, this approach is readily applicable to other lobes and clinical scenarios where serial imaging is available.\\u003c/p\\u003e\\u003cp\\u003eIt is important to consider a few clinical implications of this volumetric approach. While subtle streak-like atelectasis (grade 1) may not lead to overt volume loss, we observed significant differences in volume metrics between grade 0\\u0026ndash;1, suggesting that early changes can also be quantitatively captured. On a different note, our model segments lung lobes with all included structures such as tumor masses or areas of consolidation. And notably, unlike atelectasis, tumor masses or consolidation from pneumonia typically do not lead to a substantial reduction in lobar volume. Therefore, the volumetric approach may offer the potential to differentiate atelectasis from other causes of increased opacity, such as pneumonia. This distinction, however, would require more detailed validation.\\u003c/p\\u003e\\u003cp\\u003eThis volumetric approach for grading atelectasis also aligns with the principles of clinically explainable AI (xAI) as it yields outputs that are directly interpretable by clinicians, radiologists, or even technicians. This is unlike classification-based models, which often function as opaque \\\"black boxes\\\". Furthermore, developing a reliable classification model for atelectasis grading would require a much larger dataset to capture the wide variability in radiographic appearances, particularly subtle findings like linear opacities in grade 1\\u0026mdash;despite the positional invariance of CNNs. Hence, this segmentation-based method offers a more interpretable means of grading atelectasis and, by focusing solely on aerated volume, provides a broadly applicable solution across clinical contexts that remains effective even with limited training data.\\u003c/p\\u003e\\u003cp\\u003eThis study offers several key strengths. Firstly, it leverages a large real-world clinical cohort for automated quantification of atelectasis, using clinically interpretable AI-derived volumetry. By validating model outputs against physician-assigned atelectasis grades, the study exemplifies clinically applied AI, in contrast to much of AI research in healthcare that is decoupled from real-world outcomes. Unlike black-box classification models, our pipeline generates outputs that can be reviewed and validated by clinicians/technicians in real time, exemplifying verifiable AI. Its utility is further enhanced by the availability of open-source scripts for verifying segmentation accuracy and allowing integration into the backend of most software. A major technical strength is the use of nnU-Net\\u0026mdash;a widely validated segmentation framework in biomedical imaging\\u0026mdash;requiring relatively lesser expertise to deploy. Additionally, the volume metrics enable atelectasis to be measured as a continuous variable, preserving more information than categorical grading allowing for more nuanced statistical modeling. Moreover, continuous outcomes also provide greater statistical power and enable smaller sample sizes in clinical studies, supporting their use as efficient and meaningful endpoints in clinical trials.\\u003c/p\\u003e\\u003cp\\u003eThis work offers several potential avenues for both research and clinical application. From a research perspective, we have already applied this framework in a retrospective study (under review) assessing whether surgical fixation (\\u0026ldquo;pexy\\u0026rdquo;) of the RML reduces postoperative RML atelectasis and torsion\\u003csup\\u003e\\u003cspan citationid=\\\"CR33\\\" class=\\\"CitationRef\\\"\\u003e33\\u003c/span\\u003e\\u003c/sup\\u003e. Beyond this, the model outputs of lobar volumes/atelectasis could be used as quantitative endpoints in clinical trials focused on ventilation strategies and other interventions to prevent atelectasis, or on respiratory biomechanics across anesthesiology, critical care, and thoracic surgery. Clinically, the lobe segmentation model could be integrated into workflows such as preoperative planning of lung surgery and tumor localization through lobar mapping along with other models of tumor segmentation. It may also serve as an objective tool for postoperative assessment of atelectasis in research or clinical applications. These applications, however, warrant validation in prospective, context-specific studies.\\u003c/p\\u003e\\u003cp\\u003eDespite the strengths of this study, several limitations must be acknowledged. While the preoperative model demonstrated strong generalizability, including external validation, caution is warranted in patients with chest wall or vertebral anomalies, and at the RUL/RML junction, where fissure incompleteness is often observed. Additionally, errors (Figs.\\u0026nbsp;\\u003cspan refid=\\\"Fig6\\\" class=\\\"InternalRef\\\"\\u003e6\\u003c/span\\u003e and \\u003cspan refid=\\\"Fig7\\\" class=\\\"InternalRef\\\"\\u003e7\\u003c/span\\u003e) may stem from inconsistencies in reference standard annotations, a persistent challenge in biomedical AI. The postoperative model lacks external validation due to the absence of publicly available datasets. Furthermore, the clinical outcome of atelectasis grading was derived from a single-center retrospective cohort, which may limit generalizability. Hence until larger, more diverse datasets become available, applying automated lobe volumes for atelectasis grading may require visual confirmation. Special consideration is needed in cases with pleural pathology or external compression, which can globally affect lobar volumes and misrepresent aeration and relative lobar volumes. Importantly, the current pipeline is designed for longitudinal comparison and cannot interpret a single CT scan in isolation. Lastly, although we demonstrated the utility of continuous volume metrics as a surrogate for atelectasis, we did not define specific thresholds for grading atelectasis\\u0026mdash;an area that warrants further investigation and validation in future studies with clinical measurements like arterial blood gases, to establish meaningful thresholds for clinical classification.\\u003c/p\\u003e\"},{\"header\":\"Conclusion\",\"content\":\"\\u003cp\\u003eWe developed and validated a deep learning\\u0026ndash;based volumetric pipeline to quantify postoperative atelectasis after right upper lobectomy. Using automated lobar segmentation and validation against physician-assigned grades of atelectasis, we show that right middle lobe volume loss serves as an objective and reproducible surrogate for differentiating clinical atelectasis severity. These findings support the broader applicability of deep learning\\u0026ndash;driven lobar segmentation and volumetry as a reliable method for grading postoperative atelectasis in both clinical and research settings, and potentially for any atelectasis across other types of applications in pulmonary medicine and thoracic surgery.\\u003c/p\\u003e\"},{\"header\":\"Declarations\",\"content\":\"\\u003cp\\u003e\\u003cstrong\\u003eCode and data availability statement\\u003c/strong\\u003e: The code and the final preoperative and postoperative model checkpoints are available on Github [https://github.com/Devanish31/atelectasis] and Figshare [10.6084/m9.figshare.29877509], respectively.\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eIRB:\\u0026nbsp;\\u003c/strong\\u003eApproved by the Stanford University School of Medicine IRB (IRB-70048).\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eAuthor contributions\\u003c/strong\\u003e: DNK – Conceptualization, Data curation, Formal analysis, Methodology, Supervision, Writing – original draft, Writing – review and editing; GMF - Conceptualization, Methodology, Writing – review and editing; NL - Conceptualization, Methodology, Writing – review and editing; LLT - Conceptualization, Methodology, Writing – review and editing; NSL - Conceptualization, Methodology, Writing – review and editing; IAE - Conceptualization, Methodology, Writing – review and editing; DZL - Conceptualization, Methodology, Writing – review and editing; LMB – Conceptualization, Methodology, Writing – review and editing; MFB - Conceptualization, Methodology, Writing – review and editing; HHG - Conceptualization, Methodology, Writing – review and editing; CPL - Conceptualization, Methodology, Writing – review and editing; JBS - Conceptualization, Methodology, Supervision, Writing – original draft, Writing – review and editing. All authors read and approved of the final manuscript.\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eDeclaration of generative AI and AI-assisted technologies\\u003c/strong\\u003e: During the preparation of this work, the author(s) utilized ChatGPT to assist with rephrasing and refining the writing in the manuscript. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the published article.\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eCompeting interests\\u003c/strong\\u003e: Joseph Shrager: Consulting - Becton Dickinson; Lungpacer; Leah M. Backhus, advisory panel member with Johnson and Johnson, AstraZeneca, Genentech/Roche, and Bristol Myers Squibb; Natalie S. Lui, consulting with Intuitive Surgical Inc. and Centese, grants from Intuitive Foundation; Other authors have nothing to declare. Curt Langlotz - This research is supported in part by the Medical Imaging and Data Resource Center; grant (75N92020D00021) from National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health and through The Advanced Research Projects Agency for Health (ARPA-H). Dr. Langlotz has received consulting fees from Sixth Street; honoraria for lectures and support for travel from Singapore Ministry of Health, McKinsey, and Philips; patent pending with GE HealthCare; former president of RSNA; on board of directors and shareholder of Bunkerhill Health (since March 31, 2019); on board of directors Sirona Medical (since February 2025) option holder of Whiterabbit.ai (since October 1, 2017); advisor and option holder of Galileo CDS (since May 1, 2019); advisor and option holder of Sirona Medical (since July 6, 2020); advisor and option holder of ADRA.ai (since September 17, 2020); advisor and option holder of Cognita (since November 13, 2024); Advisor and shareholder, TurboRadiology (since January 17, 2025); and gifts to institution, department, and/or research center from BunkerHill Health, Carestream, CARPL.ai, Clairity, GE HealthCare, Google Cloud, IBM, Kheiron, Lambda, Lunit, Microsoft, Nightingale Open Science, Philips, Siemens Healthineers, Stability.ai, Subtle Medical, VinBrain, Visiana, Whiterabbit.ai, Lowenstein Foundation, and Gordon and Betty Moore Foundation.\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eFunding\\u003c/strong\\u003e: None\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eAcknowledgements\\u003c/strong\\u003e: \\u003cem\\u003eSome of\\u0026nbsp;the computing\\u0026nbsp;for this project\\u0026nbsp;was performed\\u0026nbsp;on the Sherlock\\u0026nbsp;cluster. We would\\u0026nbsp;like to thank\\u0026nbsp;Stanford University\\u0026nbsp;and the Stanford\\u0026nbsp;Research Computing\\u0026nbsp;Center for providing\\u0026nbsp;computational\\u0026nbsp;resources and\\u0026nbsp;support that\\u0026nbsp;contributed to\\u0026nbsp;these research\\u0026nbsp;results. This research used data/services provided by STARR (STAnford medicine Research data Repository), a clinical data warehouse containing patients' health information of the Stanford hospitals made possible by Stanford School of Medicine Research Office and the Stanford Clinical and Translational Science Award Number UL1TR003142 from the National Center for Advancing Translational Sciences. Data from the Stanford customized version of The Society of Thoracic Surgery Database was also used for creating this study cohort.\\u003c/em\\u003e\\u0026nbsp;\\u003c/p\\u003e\"},{\"header\":\"References\",\"content\":\"\\u003col\\u003e\\n\\u003cli\\u003eZeng C, Lagier D, Lee JW, Vidal Melo MF. Perioperative Pulmonary Atelectasis: Part I. Biology and Mechanisms. \\u003cem\\u003eAnesthesiology\\u003c/em\\u003e. 2022;136(1):181-205. \\u003c/li\\u003e\\n\\u003cli\\u003eTanner TG, Colvin MO. Pulmonary Complications of Cardiac Surgery. \\u003cem\\u003eLung\\u003c/em\\u003e. 2020;198(6):889-896. \\u003c/li\\u003e\\n\\u003cli\\u003eZhao Y, Zheng R, Xiang W, Ning D, Li Z. Systematic review and meta-analysis on perioperative intervention to prevent postoperative atelectasis complications after thoracic surgery. \\u003cem\\u003eAnn Palliat Med\\u003c/em\\u003e. 2021;10(10):107260734-107210734. \\u003c/li\\u003e\\n\\u003cli\\u003eKo E, Yoo KY, Lim CH, Jun S, Lee K, Kim YH. Is atelectasis related to the development of postoperative pneumonia? a retrospective single center study. \\u003cem\\u003eBMC Anesthesiol\\u003c/em\\u003e. 2023;23:77. \\u003c/li\\u003e\\n\\u003cli\\u003eNewman B, Krane EJ, Gawande R, Holmes TH, Robinson TE. Chest CT in children: anesthesia and atelectasis. \\u003cem\\u003ePediatr Radiol\\u003c/em\\u003e. 2014;44(2):164-172. \\u003c/li\\u003e\\n\\u003cli\\u003eKhan A, Bashour S, Sabath B, et al. Severity of Atelectasis during Bronchoscopy: Descriptions of a New Grading System (AtelectasisSeverity Scoring System-\\u0026quot;ASSESS\\u0026quot;) and At-Risk-Lung Zones. \\u003cem\\u003eDiagn Basel Switz\\u003c/em\\u003e. 2024;14(2):197. \\u003c/li\\u003e\\n\\u003cli\\u003eMcGuinness S, Parke RL. A New System For Assessing Atelectasis On Chest X-Ray After Sternotomy For Cardiac Surgery. In: \\u003cem\\u003eA48. CLINICAL RESEARCH\\u003c/em\\u003e. American Thoracic Society International Conference Abstracts. American Thoracic Society; 2011:A1679-A1679. \\u003c/li\\u003e\\n\\u003cli\\u003eAuctores. A New System for Outcome Measure of Atelectasis in Studies of Cardiac Surgery. Auctores. Accessed June 27, 2025. https://www.auctoresonline.org/article/a-new-system-for-outcome-measure-of-atelectasis-in-studies-of-cardiac-surgery\\u003c/li\\u003e\\n\\u003cli\\u003eSargent MA, McEachern AM, Jamieson DH, Kahwaji R. Atelectasis on pediatric chest CT: comparison of sedation techniques. \\u003cem\\u003ePediatr Radiol\\u003c/em\\u003e. 1999;29(7):509-513. \\u003c/li\\u003e\\n\\u003cli\\u003eTenling A, Hachenberg T, Tyd\\u0026eacute;n H, Wegenius G, Hedenstierna G. Atelectasis and gas exchange after cardiac surgery. \\u003cem\\u003eAnesthesiology\\u003c/em\\u003e. 1998;89(2):371-378. \\u003c/li\\u003e\\n\\u003cli\\u003eReber A, Nylund U, Hedenstierna G. Position and shape of the diaphragm: implications for atelectasis formation. \\u003cem\\u003eAnaesthesia\\u003c/em\\u003e. 1998;53(11):1054-1061. \\u003c/li\\u003e\\n\\u003cli\\u003eWarner DO, Warner MA, Ritman EL. Atelectasis and chest wall shape during halothane anesthesia. \\u003cem\\u003eAnesthesiology\\u003c/em\\u003e. 1996;85(1):49-59. \\u003c/li\\u003e\\n\\u003cli\\u003ePieters ALP, van der Veer T, Meerburg JJ, et al. Structural Lung Disease and Clinical Phenotype in Bronchiectasis Patients: The EMBARC CT Study. \\u003cem\\u003eAm J Respir Crit Care Med\\u003c/em\\u003e. 2024;210(1):87-96. \\u003c/li\\u003e\\n\\u003cli\\u003eRosenow T, Oudraad MCJ, Murray CP, et al. PRAGMA-CF. A Quantitative Structural Lung Disease Computed Tomography Outcome in Young Children with Cystic Fibrosis. \\u003cem\\u003eAm J Respir Crit Care Med\\u003c/em\\u003e. 2015;191(10):1158-1165. \\u003c/li\\u003e\\n\\u003cli\\u003eLevine BE, Johnson RP. Effects of atelectasis on pulmonary surfactant and quasi-static lung mechanics. \\u003cem\\u003eJ Appl Physiol\\u003c/em\\u003e. 1965;20(5):859-864. \\u003c/li\\u003e\\n\\u003cli\\u003eNoll E, Ohana M, Hengen M, et al. Validation of MRI for Volumetric Quantification of Atelectasis in the Perioperative Period: An Experimental Study in Swine. \\u003cem\\u003eFront Physiol\\u003c/em\\u003e. 2019;10:695. \\u003c/li\\u003e\\n\\u003cli\\u003eDavid M, Karmrodt J, Bletz C, et al. Analysis of atelectasis, ventilated, and hyperinflated lung during mechanical ventilation by dynamic CT. \\u003cem\\u003eChest\\u003c/em\\u003e. 2005;128(5):3757-3770. \\u003c/li\\u003e\\n\\u003cli\\u003eMizukami Y, Takahashi Y, Maki R, Adachi H. Risk factors for atelectasis of the middle lobe after right upper lobectomy: preoperative bronchial diameter and stapling of the fissure. \\u003cem\\u003eJ Thorac Dis\\u003c/em\\u003e. 2021;13(10):5649-5657. \\u003c/li\\u003e\\n\\u003cli\\u003eTejani AS, Klontzas ME, Gatti AA, et al. Checklist for Artificial Intelligence in Medical Imaging (CLAIM): 2024 Update. \\u003cem\\u003eRadiol Artif Intell\\u003c/em\\u003e. 2024;6(4):e240300. \\u003c/li\\u003e\\n\\u003cli\\u003eD Slicer image computing platform. 3D Slicer. Accessed July 26, 2025. https://slicer.org/\\u003c/li\\u003e\\n\\u003cli\\u003eAmini E, Klein R. Lung Lobe Segmentation and Quantification Dataset. Published online July 8, 2024. Accessed June 14, 2025. https://zenodo.org/records/12690803\\u003c/li\\u003e\\n\\u003cli\\u003eIsensee F, Petersen J, Klein A, et al. nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation. arXiv:1809.10486.\\u003c/li\\u003e\\n\\u003cli\\u003eIsensee F, Wald T, Ulrich C, et al. nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation. arXiv:2404.09556. \\u003c/li\\u003e\\n\\u003cli\\u003eLOLA11 - Grand Challenge. grand-challenge.org. Accessed August 21, 2025. https://lola11.grand-challenge.org/Details/\\u003c/li\\u003e\\n\\u003cli\\u003eBragman FJS, McClelland JR, Jacob J, Hurst JR, Hawkes DJ. Pulmonary Lobe Segmentation with Probabilistic Segmentation of the Fissures and a Groupwise Fissure Prior. \\u003cem\\u003eIEEE Trans Med Imaging\\u003c/em\\u003e. 2017;36(8):1650-1663. \\u003c/li\\u003e\\n\\u003cli\\u003eImran AAZ, Hatamizadeh ,Ali, Ananth ,Shilpa P., Ding ,Xiaowei, Tajbakhsh ,Nima, and Terzopoulos D. Fast and automatic segmentation of pulmonary lobes from chest CT using a progressive dense V-network. \\u003cem\\u003eComput Methods Biomech Biomed Eng Imaging Vis\\u003c/em\\u003e. 2020;8(5):509-518. \\u003c/li\\u003e\\n\\u003cli\\u003eXie W, Jacobs C, Charbonnier JP, van Ginneken B. Relational Modeling for Robust and Efficient Pulmonary Lobe Segmentation in CT Scans. \\u003cem\\u003eIEEE Trans Med Imaging\\u003c/em\\u003e. 2020;39(8):2664-2675. \\u003c/li\\u003e\\n\\u003cli\\u003ePeng Y, Zhang J. Lung lobe segmentation in computed tomography images based on multi-feature fusion and ensemble learning framework. \\u003cem\\u003eInt J Imaging Syst Technol\\u003c/em\\u003e. 2023;33(6):2088-2099. \\u003c/li\\u003e\\n\\u003cli\\u003ePang H, Wu Y, Qi S, et al. A fully automatic segmentation pipeline of pulmonary lobes before and after lobectomy from computed tomography images. \\u003cem\\u003eComput Biol Med\\u003c/em\\u003e. 2022;147:105792. \\u003c/li\\u003e\\n\\u003cli\\u003eMalik H, Anees T, Al-Shamaylehs AS, Alharthi SZ, Khalil W, Akhunzada A. Deep Learning-Based Classification of Chest Diseases Using X-rays, CT Scans, and Cough Sound Images. \\u003cem\\u003eDiagn Basel Switz\\u003c/em\\u003e. 2023;13(17):2772. \\u003c/li\\u003e\\n\\u003cli\\u003eMalik H, Anees T. Multi-modal deep learning methods for classification of chest diseases using different medical imaging and cough sounds. \\u003cem\\u003ePloS One\\u003c/em\\u003e. 2024;19(3):e0296352. \\u003c/li\\u003e\\n\\u003cli\\u003eRajpurkar P, Irvin J, Ball RL, et al. Deep learning for chest radiograph diagnosis: A retrospective comparison of the CheXNeXt algorithm to practicing radiologists. \\u003cem\\u003ePLOS Med\\u003c/em\\u003e. 2018;15(11):e1002686. \\u003c/li\\u003e\\n\\u003cli\\u003eKamtam DN, Liou DZ, Berry MF, et al. Does fixation of the right middle lobe/lingula during right upper lobectomy/left apical trisegmentectomy reduce rates of postoperative torsion/atelectasis? https://doi.org/10.1101/2025.07.28.25332308. \\u003c/li\\u003e\\n\\u003cli\\u003eDoel T, Matin TN, Gleeson FV, Gavaghan DJ, Grau V. Pulmonary lobe segmentation from CT images using fissureness, airways, vessels and multilevel B-splines. In: \\u003cem\\u003e2012 9th IEEE International Symposium on Biomedical Imaging (ISBI)\\u003c/em\\u003e. ; 2012:1491-1494. \\u003c/li\\u003e\\n\\u003cli\\u003eD U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation | SpringerLink. Accessed June 26, 2025. https://link.springer.com/chapter/10.1007/978-3-319-46723-8_49\\u003c/li\\u003e\\n\\u003cli\\u003ePark J, Yun J, Kim N, et al. Fully Automated Lung Lobe Segmentation in Volumetric Chest CT with 3D U-Net: Validation with Intra- and Extra-Datasets. \\u003cem\\u003eJ Digit Imaging\\u003c/em\\u003e. 2020;33(1):221-230. \\u003c/li\\u003e\\n\\u003cli\\u003eLassen-Schmidt B, Hering A, Krass S, Meine H. Automatic segmentation of the pulmonary lobes with a 3D u-net and optimized loss function. arXiv:2006.00083. \\u003c/li\\u003e\\n\\u003cli\\u003eFerreira FT, Sousa P, Galdran A, Sousa MR, Campilho A. End-to-End Supervised Lung Lobe Segmentation. In: \\u003cem\\u003e2018 International Joint Conference on Neural Networks (IJCNN)\\u003c/em\\u003e. ; 2018:1-8. \\u003c/li\\u003e\\n\\u003cli\\u003eGerard SE, Reinhardt JM. Pulmonary Lobe Segmentation Using A Sequence of Convolutional Neural Networks For Marginal Learning. In: \\u003cem\\u003e2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019)\\u003c/em\\u003e. ; 2019:1207-1211. \\u003c/li\\u003e\\n\\u003cli\\u003eZheng S, Nie W, Pan L, et al. A dual-attention V-network for pulmonary lobe segmentation in CT scans. \\u003cem\\u003eIET Image Process\\u003c/em\\u003e. 2021;15(8):1644-1654. \\u003c/li\\u003e\\n\\u003cli\\u003eZhang Z, Ren J, Tao X, et al. Automatic segmentation of pulmonary lobes on low-dose computed tomography using deep learning. \\u003cem\\u003eAnn Transl Med\\u003c/em\\u003e. 2021;9(4):291. \\u003c/li\\u003e\\n\\u003cli\\u003eBao N, Yuan Y, Luo Q, Li Q, Zhang LB. Edge-enhancement cascaded network for lung lobe segmentation based on CT images. \\u003cem\\u003eFront Phys\\u003c/em\\u003e. 2023;11. \\u003c/li\\u003e\\n\\u003cli\\u003eNomura GRO, Luong AT, Prakash A, et al. TriSwinUNETR lobe segmentation model for computing DIR-free CT-ventilation. \\u003cem\\u003eFront Oncol\\u003c/em\\u003e. 2025;15:1475133. \\u003c/li\\u003e\\n\\u003cli\\u003evan Rikxoort EM, Prokop M, de Hoop B, Viergever MA, Pluim JPW, van Ginneken B. Automatic segmentation of pulmonary lobes robust against incomplete fissures. \\u003cem\\u003eIEEE Trans Med Imaging\\u003c/em\\u003e. 2010;29(6):1286-1296. \\u003c/li\\u003e\\n\\u003cli\\u003eLassen B, van Rikxoort EM, Schmidt M, Kerkstra S, van Ginneken B, Kuhnigk JM. Automatic segmentation of the pulmonary lobes from chest CT scans based on fissures, vessels, and bronchi. \\u003cem\\u003eIEEE Trans Med Imaging\\u003c/em\\u003e. 2013;32(2):210-222. \\u003c/li\\u003e\\n\\u003c/ol\\u003e\"}],\"fulltextSource\":\"\",\"fullText\":\"\",\"funders\":[],\"hasAdminPriorityOnWorkflow\":false,\"hasManuscriptDocX\":true,\"hasOptedInToPreprint\":true,\"hasPassedJournalQc\":\"\",\"hasAnyPriority\":true,\"hideJournal\":false,\"highlight\":\"\",\"institution\":\"\",\"isAcceptedByJournal\":true,\"isAuthorSuppliedPdf\":false,\"isDeskRejected\":\"\",\"isHiddenFromSearch\":false,\"isInQc\":false,\"isInWorkflow\":false,\"isPdf\":false,\"isPdfUpToDate\":true,\"isWithdrawnOrRetracted\":false,\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"npj-digital-medicine\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":false,\"externalIdentity\":\"npjdigitalmed\",\"sideBox\":\"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)\",\"snPcode\":\"41746\",\"submissionUrl\":\"https://submission.springernature.com/new-submission/41746/3\",\"title\":\"npj Digital Medicine\",\"twitterHandle\":\"\",\"acdcEnabled\":true,\"dfaEnabled\":true,\"editorialSystem\":\"stoa\",\"reportingPortfolio\":\"NPJ\",\"inReviewEnabled\":true,\"inReviewRevisionsEnabled\":true},\"keywords\":\"artificial intelligence, deep learning, lobectomy, middle lobe, atelectasis, segmentation, volumetry, nnU-Net\",\"lastPublishedDoi\":\"10.21203/rs.3.rs-7768040/v1\",\"lastPublishedDoiUrl\":\"https://doi.org/10.21203/rs.3.rs-7768040/v1\",\"license\":{\"name\":\"CC BY 4.0\",\"url\":\"https://creativecommons.org/licenses/by/4.0/\"},\"manuscriptAbstract\":\"\\u003ch2\\u003eObjective:\\u003c/h2\\u003e\\u003cp\\u003eChronic postoperative atelectasis of lung lobes is an occasional complication following thoracic surgery, particularly middle lobe atelectasis following right upper lobectomy (RULobectomy). Existing methods of grading atelectasis are typically manual, subjective, and not scalable. We aimed to develop and validate an automated, deep learning\\u0026ndash;based volumetric framework to quantify and grade postoperative atelectasis using pre- and post-operative CT scans.\\u003c/p\\u003e\\u003ch2\\u003eMethods:\\u003c/h2\\u003e\\u003cp\\u003eWe retrospectively included all patients who underwent RULobectomy in our institution from 2008 to 2023 who had available pre-operative and 6-month postoperative CT scans. We trained two separate nnU-Net v2 segmentation models for preoperative and postoperative lobar anatomy followed by volumetric quantification of the right middle lobe (RML), right lower lobe (RLL), and total lung volume. Atelectasis severity in the RML was independently graded by two surgeons using a standardized, 5-point radiological scale (none, minimal, subsegmental, segmental, lobar). The association between volume metrics and clinical atelectasis severity was evaluated using both the original 5-point scale and a pooled 3-point scale (none, minimal\\u0026ndash;subsegmental, segmental\\u0026ndash;lobar).\\u003c/p\\u003e\\u003ch2\\u003eResults:\\u003c/h2\\u003e\\u003cp\\u003e236 patients comprised the study cohort. The pre- and postoperative models achieved high segmentation accuracy (mean Dice scores: 0.98\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.02 and 0.99\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.00, respectively). Median (IQR) RML volume loss progressively increased with higher atelectasis grades, from \\u0026minus;\\u0026thinsp;4.6 mL (-78.5, 59.0) in grade 0 to \\u0026minus;\\u0026thinsp;317.8 mL (-440.7, -194.8) in grade 4 atelectasis (p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.001). Normalized RML/right lung (RL) and RML/total lung (TL) volume ratios showed statistically significant differences across the pooled atelectasis grades (p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.001). Conversely, normalized RLL volumes increased with worsening RML atelectasis (p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.001), suggesting compensatory hyperinflation.\\u003c/p\\u003e\\u003ch2\\u003eConclusions:\\u003c/h2\\u003e\\u003cp\\u003eWe demonstrate the feasibility and clinical relevance of deep learning\\u0026ndash;based volumetric assessment of atelectasis after RULobectomy. This automated pipeline with its open-source model enables reproducible quantification of atelectasis through lobar volume loss and may serve as a scalable tool for clinical and research applications involving atelectasis.\\u003c/p\\u003e\",\"manuscriptTitle\":\"Deep-learning based quantitative evaluation of postoperative atelectasis following right upper lobectomy\",\"msid\":\"\",\"msnumber\":\"\",\"nonDraftVersions\":[{\"code\":1,\"date\":\"2025-10-08 09:18:16\",\"doi\":\"10.21203/rs.3.rs-7768040/v1\",\"editorialEvents\":[{\"type\":\"communityComments\",\"content\":0},{\"type\":\"decision\",\"content\":\"Revision requested\",\"date\":\"2025-11-04T01:01:27+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"90606757820033088369093618706234620869\",\"date\":\"2025-11-03T09:01:43+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"136975970004653235211193524545653537594\",\"date\":\"2025-11-02T23:50:04+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"editorInvitedReview\",\"content\":\"\",\"date\":\"2025-11-01T07:45:50+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"204116608675267216547531958724166047817\",\"date\":\"2025-11-01T01:20:20+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"editorInvitedReview\",\"content\":\"\",\"date\":\"2025-10-31T23:58:41+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"112535925878800316775520379777748188983\",\"date\":\"2025-10-31T21:30:41+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"editorInvitedReview\",\"content\":\"\",\"date\":\"2025-10-20T21:08:55+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"74483921672650740201924940823550141350\",\"date\":\"2025-10-20T08:12:14+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"132170043108016544316243419531408754962\",\"date\":\"2025-10-18T23:45:11+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"97373360239871548798248727176867438321\",\"date\":\"2025-10-16T07:41:06+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewersInvited\",\"content\":\"\",\"date\":\"2025-10-16T04:10:17+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"editorAssigned\",\"content\":\"\",\"date\":\"2025-10-13T23:23:11+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"checksComplete\",\"content\":\"\",\"date\":\"2025-10-13T07:29:37+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"submitted\",\"content\":\"npj Digital Medicine\",\"date\":\"2025-10-02T16:35:57+00:00\",\"index\":\"\",\"fulltext\":\"\"}],\"status\":\"published\",\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"npj-digital-medicine\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":false,\"externalIdentity\":\"npjdigitalmed\",\"sideBox\":\"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)\",\"snPcode\":\"41746\",\"submissionUrl\":\"https://submission.springernature.com/new-submission/41746/3\",\"title\":\"npj Digital Medicine\",\"twitterHandle\":\"\",\"acdcEnabled\":true,\"dfaEnabled\":true,\"editorialSystem\":\"stoa\",\"reportingPortfolio\":\"NPJ\",\"inReviewEnabled\":true,\"inReviewRevisionsEnabled\":true}}],\"origin\":\"\",\"ownerIdentity\":\"963e029c-706b-4158-8f8f-6d19f8af4e1e\",\"owner\":[],\"postedDate\":\"October 8th, 2025\",\"published\":true,\"recentEditorialEvents\":[],\"rejectedJournal\":[],\"revision\":\"\",\"amendment\":\"\",\"status\":\"published-in-journal\",\"subjectAreas\":[{\"id\":55863034,\"name\":\"Health sciences/Diseases\"},{\"id\":55863035,\"name\":\"Health sciences/Medical research\"}],\"tags\":[],\"updatedAt\":\"2026-05-04T15:59:39+00:00\",\"versionOfRecord\":{\"articleIdentity\":\"rs-7768040\",\"link\":\"https://doi.org/10.1038/s41746-026-02683-6\",\"journal\":{\"identity\":\"npj-digital-medicine\",\"isVorOnly\":false,\"title\":\"npj Digital Medicine\"},\"publishedOn\":\"2026-04-30 15:57:09\",\"publishedOnDateReadable\":\"April 30th, 2026\"},\"versionCreatedAt\":\"2025-10-08 09:18:16\",\"video\":\"\",\"vorDoi\":\"10.1038/s41746-026-02683-6\",\"vorDoiUrl\":\"https://doi.org/10.1038/s41746-026-02683-6\",\"workflowStages\":[]},\"version\":\"v1\",\"identity\":\"rs-7768040\",\"journalConfig\":\"researchsquare\"},\"__N_SSP\":true},\"page\":\"/article/[identity]/[[...version]]\",\"query\":{\"redirect\":\"/article/rs-7768040\",\"identity\":\"rs-7768040\",\"version\":[\"v1\"]},\"buildId\":\"XKTyCvWXoU3ODBz1xrDgd\",\"isFallback\":false,\"isExperimentalCompile\":false,\"dynamicIds\":[84888],\"gssp\":true,\"scriptLoader\":[]}","source_license":"CC-BY-4.0","license_restricted":false}