Global and regional accuracy of deep learning-based tumor segmentation from whole-body [¹⁸F]fluorodeoxyglucose PET/CT images

Global and regional accuracy of deep learning-based tumor segmentation from whole-body [¹⁸F]fluorodeoxyglucose PET/CT images | 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 Research Article Global and regional accuracy of deep learning-based tumor segmentation from whole-body [¹⁸F]fluorodeoxyglucose PET/CT images Andrea Ciarmiello, Nikola Yosifov, Donatella Masciale, Ornella Ferrando, and 8 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6895938/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 28 Nov, 2025 Read the published version in EJNMMI Research → Version 1 posted 5 You are reading this latest preprint version Abstract Background The number of [¹⁸F]fluorodeoxyglucose ([¹⁸F]FDG)-PET/CT scans performed has significantly increased in the last decade in line with the increasing trend of oncological malignancies. Such images, which signal high glucose-uptake areas are key in defining the extent of the disease, staging and response to therapy. Processing and evaluation of ([¹⁸F]FDG)-PET/CT scans, however, require manual annotation by well-trained specialists and above all time. In time and resource-constrained settings meeting the increasing demand for PET/CT scans has become challenging. The main goal of our study was to test the relationship between the volumes predicted by the deep learning algorithm and the manually segmented ones. The secondary objective goal was to measure the extent at which the predictive accuracy is associated with normal background uptake. Results The study sample included 1159 [¹⁸F]FDG-PET/CT scans from subjects with histologically confirmed diagnoses of lung cancer, lymphoma, and melanoma. 881 (70%) [¹⁸F]FDG-PET/CT scans were used as the training dataset and 232 (20%) scans were used as an internal validation dataset. A subsample of 116 (10%) [¹⁸F]FDG-PET/CT scans not used for training was used as the test dataset. The segmentation model was implemented with the nnU-Net convolutional network available in the MONAI framework. Model performance was measured with the Dice score. Correlation between manual and predicted segmentation was assessed using linear correlation. Totalsegmentator tool was used to identify lesions location and assess the tumor-to-background ratio (TBR) for quantitative analysis. Network achieved Dice scores of 0.805 (validation) and 0.784 (test), showing strong agreement with manual segmentations. Anatomical localization was successful in 74% of the 7914 detected lesions. High correlation (R=0.88, p2 had significantly better Dice scores than those with lower contrast (TBR ≤ 1–2 or ≤1). Conclusions These results are consistent with previous reports on PET-based segmentation, further validating nnU-Net as a reliable approach for detecting hypermetabolic lesions and assessing global disease burden in FDG-PET imaging. Moreover, the significant relationship between TBR and segmentation accuracy suggests the possibility of further improvements by integrating metabolic profile into the predictive model. Deep learning nnU-Net Segmentation 18F-FDG PET/CT Figures Figure 1 Figure 2 Figure 3 Introduction Staging of tumor lesions and the monitoring of response to many cancer therapy regimens including immune checkpoint inhibitors is based on [¹⁸F]fluorodeoxyglucose ([¹⁸F]FDG)-PET/CT images signaling areas of high glucose uptake[ 1 , 2 ]. The lesions are accurately measured to establish the patient’s overall disease burden, by means of manual segmentation performed by a highly trained specialists (radiologist or nuclear medicine physician). The process is painstaking and may require up to 5 minutes for a single lesion and several hours in patients with multiple lesions [ 3 ], thus occupying a large portion of the workload of nuclear medicine physicians. With the increasing trend of oncological malignancies and increase in the number of PET/CT scans estimated to be the 13% in the last decade alone [ 4 – 7 ] several efforts have been made to alleviate this burden with the development in advanced imaging technology and non-invasive diagnostic techniques [ 8 , 9 ]. Nonetheless, many tasks involved in staging and follow-up remain operator-dependent and as such are subject to errors [ 10 , 11 ], and inter- and intra-observer variability [ 12 ]. Recently, the introduction of several deep learning algorithms have provided additional tools for quicker and more accurate tumor assessment to support patient-tailored treatments [ 13 – 15 ]. Artificial intelligence (AI) with deep neural network (DNN) frameworks has already been used for segmentation of normal anatomical structures using computed tomography (CT) images [ 16 , 17 ]. To date, the most convincing open-source software is that offered by Wasserthal et al. [ 17 ] where a DNN model is able to segment 117 anatomical structures. AI-based DNNs models have also been proposed to automatically segment hypermetabolic malignant lesions from [¹⁸F]FDG-PET/CT images [ 18 – 21 ] or to correctly classify hypermetabolic regions of non-tumor nature as in the case of physiologically hypermetabolic tissues or high uptake due to infectious and inflammatory diseases [ 22 – 24 ]. This task is challenging as in the first case, the high metabolic activity of the healthy background tissue could limit the model's ability to recognize the tumor lesion, while in the latter case the model could classify a benign pathological process as malignant. In order to evaluate the global and anatomical region-specific performance of the convolutional neural network segmentation model, we retrospectively studied its diagnostic accuracy using a sample of 3D [¹⁸F]FDG-PET/CT scans of patients diagnosed with lymphoma, melanoma, and lung cancer. The primary objective of our study was to test the relationship between the neural network predicted and manually segmented volumes. The secondary objective was to measure the extent at which the predictive accuracy is associated with normal background uptake. Materials and Methods Dataset This retrospective study was carried on FDG-PET/CT images from the autoPET dataset [ 25 ] and the LOMICS20 dataset [ 15 ]. Both studies have been approved by the respective regional ethics committees. autoPET dataset This dataset includes FDG-PET/CT scans of 145 lymphomas, 168 lung cancers, and 188 melanomas, as well as a 513 scans of negative controls. Negative PET/CT scans came from patients studied after medical or surgical treatment with no evidence of metabolically active disease. Image voxel size is (2.04 x 2.04 x 3.00) mm³. 3D volumes of manual annotations are provided for all PET/CT scan. Manual annotation was made by two experienced radiologists with ten and five years of experience. autoPet cohort is publicly available at TCIA (The Cancer Imaging Archive) [ 26 ]. LOMICS20 dataset A subset of images from the LOMICS20 cohort consisting of 145 whole-body FDG-PET/CT scans of lung cancer patients. Image voxel size is (2.73 x 2.73 x 3.27) mm³. 3D volumes of manual annotations provided for all PET/CT scan was performed by two board-certified nuclear medicine physicians with five years of experience. LOMICS20 Dataset is not publicly available. In order to train, validate and test the prediction model on different samples, the cohort was randomly split into three datasets. Therefore, a total of 881 [¹⁸F]FDG-PET/CT scans (70% of the entire dataset) were used as the training set, and 232 scans (20% of the dataset) were used for internal validation. Additionally, a subsample of 116 scans (10% of the dataset), which were not part of the training or validation sets, was designated as the test dataset (Table 1). Random splitting was performed with the sklearn platform in Python 3 to obtain datasets comparable by age, gender, diagnosis and cases with hypermetabolic lesions. Procedures PET/CT acquisition and processing was performed in agreement with international guidelines for PET/CT examinations in oncology [ 27 ]. The radiotracer [¹⁸F]FDG was administered after 6 hours of fasting and blood glucose level check in all patients. Scans were performed at least 60 minutes after its administration. Details on acquisition and image reconstruction parameters are reported elsewhere [ 15 , 25 ]. As by clinical standards, PET/CT images were classified as positive in the presence of one or more lesions featuring high [¹⁸F]FDG uptake. The positive sample consisted of 188 melanomas, 313 lung tumors, and 145 lymphomas. The PET/CT images were reviewed by board-certified nuclear medicine physicians who manually segmented positive lesions using dedicated software, NORA image analysis platform (University of Freiburg, Germany), or PET Volume Computerized Assisted Reporting (PETVCAR) –a commercial software running on Advantage Workstation (version 4.6; GE Healthcare). The resulting 3D binary mask was used as ground-truth in the training, validation and testing phases. Network model The segmentation model was implemented in the nnU-Net framework. This is a U-Net based network able of configuring all hyperparameters on the dataset feature [ 28 , 29 ]. This architecture was selected for its reported high performance in segmenting both normal and pathological 3D medical images from different imaging modalities [ 17 , 30 , 31 ] The network architecture consisting of six stages in the encoder, with (1, 3, 4, 6, 6, 6) blocks in each stage, and (16, 32, 64, 128, 256, 320) feature layers per stage, respectively and the Leaky ReLU activation function.The network parameters were optimized over 1000 epochs using the Adam optimizer, with a learning rate of 1e-4, a weight decay of 1e-5, a dropout rate of 0.20, and a batch size of 2. Patches of (96×128×128) voxels were extracted through a sliding window approach with an overlap of 0.25 between consecutive patches [ 32 ]. To avoid overfitting, an early stopping strategy with a patience of 60 epochs and a delta of 0.01 is implemented. Preprocessing and data augumentation CT images were downsampled to match the corresponding PET image size. PET, CT, and label images were then resampled to a common imaging resolution, with uniform isotropic spacing of 2 mm³ using bilinear interpolation for PET and CT images and nearest-neighbor interpolation for label images. PET images were then processed by applying Min-Max normalization to standardized uptake values ​​(SUV) [ 33 ]. To augment the input to the network we applied at start of every epoch a series of random transformations to PET and label images including a) 3D elastic deformations applying a Gaussian kernel with standard deviation and offsets uniformly sampled from (0, 1), b) random scaling by 1.1 factor along axes, c) axial rotations by (−π/12, π/12) angle, d) 3D translations in the range (0, 10) voxels along axes, e) addition of random Gaussian noise (µ = 0 ,σ = 1) and f) Gamma correction (γ ∈ 0.7, 1.5) [ 32 , 33 ] Lesions anatomical mapping TotalSegmentator was used to segment anatomical structures from PET/CT images to localize hypermetabolic lesions and assess tumor-to-background ratio (TBR) for quantitative analysis ( https://github.com/wasserth/TotalSegmentator ) [ 17 ]. Lesions were mapped onto the corresponding anatomical structure by matching the lesion center of mass with the spatial coordinates of the anatomical map. Lesions whose center of mass did not match the coordinates of any extracted tissue were used for global assessment only and not included in the regional analysis. Lesions mapping was performed with Scipy ndimage library ( https://github.com/scipy/scipy ) [ 34 ]. SUV estimate SUV images were used to calculate metabolic tumor volume (MTV) and total lesion glycolysis (TLG). Tumor-to-background ratio (TBR) was calculated by dividing the lesion SUVmax by the SUVmean of corresponding background tissue. Background activity was measured on the lesion's anatomical structure after masking the region corresponding to the hypermetabolic area as defined by annotated label. Development environment The project was carried out using the open-source platform MONAI [ 35 ] ( https:/monai.io ), which is designed to standardize medical image processing by means of a broad selection of AI-based processing tools including image pre-processing, training, and post-processing operations needed in the formulation of prediction models. Image preprocessing and network implementation were performed in Python version 3.11.5 and cuda 12.2, using Python-based image processing libraries, such as Scikit-image version 0.20.0 ( https://scikit-image.org/ ), OpenCV version 4.6.0 ( https://opencv.org/ ), and SciPy version 1.11.3 [ 34 ] ( https://scipy.org/ ). Python scripts to convert images from DICOM to NIFTI and to generate CT volume resampled to PET, and standardized uptake values (SUV) images were available at https://github.com/lab-midas/TCIA processing. All experiments were performed under Ubuntu 22.04.3 using a Windows 11 Pro 23H2 64-bit workstation with 128GB of RAM, a 2.40GHz, 24-core Intel Xeon Silver 4214R CPU. Processing-intensive tasks were handled by an NVIDIA RTX A4000 graphic card equipped with 16 GB RAM. Statistics Statistical analysis was performed with RStudio 4.4.2 with the R libraries dplyr, tidyr, stringr, gtsummary, gt and tidyverse. Data were reported as median with range, unless otherwise specified. Categorical variables were reported as counts and percentages. Group comparisons used One-way analysis and Chi-square test for categorical variables. Correlations between manual and predicted segmentation were assessed using linear correlation. Spearman's rho was used to estimate the strength of association between prediction accuracy and TBR. The network was optimized using a Dice loss function [ 36 ], with model performance quantified by the Dice Similarity Coefficient (DSC). During training, validation was performed at every epoch, with model selection based on maximal DSC achievement. Model classification performance was evaluated using a confusion matrix (CM), comprising true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). From the CM Sensitivity (SS), Specificity (SP), Positive predictive value (PPV), Negative predictive value (NPV), and Balanced accuracy (BA) were derived. Performance was assessed across training, validation, and test datasets for PET/CT scans with and without hypermetabolic tumor sites. A p-value less than 0.05 was considered statistically significant. Results Clinical and demographic characteristics of the study cohort are reported in Table 1. No differences were found in the distribution of demographic characteristics, diagnosis and positive cases number of the training, validation and test datasets (p > 0.05). Among 1159 [¹⁸F]FDG-PET/CT scans in the study population, 646 (55.7%) showed positive findings for hypermetabolic lesions, with 7914 distinct high-glucose-metabolism lesions identified overall. The PET/CT segmentation model exhibited progressive convergence, with Dice loss decreasing consistently over 1000 training epochs (Supplementary Fig. S1 ). At completion, the model achieved a peak aggregate Dice similarity coefficient (DSC) of 0.805 on the validation set. No significant differences were found between the performance of the predictive model in the validation and test datasets (P = 0.13). Density distribution of the dice score in the two groups is reported in Fig. 1 , panel A. In the [¹⁸F]FDG-positive cases, the manually annotated lesion volume significantly correlated with the predicted lesion volume ( R 2 = 0.88; p < 0.0001). Figure 1 , panel B shows the plot of the ground-truth volume versus the predicted volume. Figure 2 shows a representative image of automatic segmentation. Panel A shows [¹⁸F]FDG-PET/CT images with an area of ​​intense uptake in the upper lobe of the right lung. Panel B shows the grayscale map of segmented anatomical structures from the PET/CT images in the background. The overlaid color map shows true positive and false negative in red and cyan, respectively. The proposed method achieved median Dice similarity coefficients of 0.70 (validation set) and 0.66 (test set) for lesion-level segmentation accuracy (Table 2). In the validation cohort, the model showed balanced accuracy of 0.82 (95% CI 0.74–0.87), with positive predictive value (PPV) of 0.77 (95% CI 0.64–0.85) and negative predictive value (NPV) of 0.97 (95% CI 0.94–0.98). Performance in the test cohort showed sustained robustness, with a balanced accuracy of 0.79 (95% CI 0.69–0.85), maintaining diagnostic reliability comparable to the validation set. The model achieved a PPV of 0.83 (95% CI 0.73–0.86), and maintained NPV 0.96 (95% CI 0.93–0.98) (Table 2). Of 7914 hypermetabolic lesions, 5862 (78%) had a center of mass matching the spatial coordinates of key anatomical structures: adrenal glands, bones, brain, esophagus, kidneys, gallbladder, large and small intestines, liver, lungs, muscles, pancreas, prostate, spinal cord, spleen, stomach, thyroid, and trachea. Performance metrics, including true positive and false negative rates, showed significant inter-regional variability (Fig. 3 , panel A). A significant positive correlation was observed between the Dice score and tumor-to-background ratio (R = 0.61, P = 0.0082). Figure 3 B shows that the strength of this correlation varied by anatomical region, with higher Dice scores consistently associated with regions exhibiting elevated TBR. This relationship was further confirmed by one-way ANOVA, which showed significant differences in Dice scores across TBR-stratified groups (≤ 1, 1–2, > 2; p < 0.001), as detailed in Table 3. Discussion In recent years, several studies have confirmed the relevance of deep learning in the recognition and quantitative measurement of [¹⁸F]FDG-avid lesions in oncology and AI-based tools for automatized segmentation of hypermetabolic lesions from FDG-PET/CT scans [ 18 , 37 , 38 ]. Some studies specifically including large populations of 5575 patients with lymphoma have reported lesion segmentation models able to achieve an accuracy of 0.875 [ 19 ]. Further studies have attempted to evaluate the accuracy of automatic segmentation on datasets including multiple oncological diseases such as Jemaa et al who reported a dice score of 0.873 on 3664 [¹⁸F]FDG-PET/CT scans [ 39 ]. So far, however, the majority of these results have been achieved from whole-body PET/CT scan datasets, where the overall lesion load is distributed over a significantly larger background, featuring significantly different radiotracer uptake across the anatomical structures included in the scan. These conditions might affect the predictive performance, resulting in differences in regional compared to global lesion accuracy. Some studies have focused on oncological disease localized in specific anatomical regions using datasets consisting of 1 or 2 bed scans thus obtaining a smaller background region and presumably more homogeneous radiotracer uptake levels. Among these, Huang et al. achieved a dice score of 0.732 on a dataset of 22 head and neck cancer patients [ 40 ] and Teramoto et al., obtained a dice score of 0.85 on a sample of 104 [¹⁸F]FDG-PET/CT of patients with solitary lung nodules [ 41 ]. Such studies, however, were based on small sample sizes, which is a potential drawback for deep learning applications. Furthermore, non-whole-body scans do not allow estimation of the global disease burden, which is one of the expected results of automatic segmentation. As previously reported, normal tissues can have significantly different levels of [¹⁸F]FDG-PET/CT uptake such as cerebrum, cerebellum, myocardium, tonsils, liver and spleen which generally have higher uptake than other tissues. Knowledge of physiological uptake is therefore required for correct interpretation of PET/CT studies [ 42 , 43 ]. Hence, our study also evaluated whether predictive performance was associated with physiological variations in normal background uptake across different tissue types. By means of the nnU-Net model implemented in the MONAI framework we were able to segment [¹⁸F]FDG-avid lesions from PET/CT scans. Model segmentation performance measured by the dice score achieved a DCSs of 0.805 and 0.784 on the validation and test datasets, respectively indicating a high overlap between the model segmented and the manually-annotated lesions. The correlation between lesion volume segmented by the model and manually-annotated lesion volumes was almost linear (R2 = 0.88, P < 0.0001) and is in line with a previous study that has reported the linear correlation between lesion volumes detected by nnU-NET model and manual annotated volumes [ 26 ]. Moreover, we found that TBR is significantly associated with prediction accuracy as measured by the dice score (TBR vs. dice score = 0.61; P = 0.0082), and that lower TBR values are associated with significant worse predictive performance. As to the AI-based assessment of global disease burden, this has been studied in previous work for several oncological diseases [ 37 , 44 – 46 ]. Disease burden quantification implies the segmentation of all hypermetabolic lesions to measure metabolically active tumor volumes, mean and max lesion activity, total metabolic tumor volume, and total lesion glycolysis [ 47 ]. Fully-automated [¹⁸F]FDG-PET/CT image segmentation models are particularly useful for staging and assessment of treatment response in several malignancies where the measurement of the overall burden of disease is essential for patient management. Among these, Häggström et al. implemented a deep learning-based model for PET/CT segmentation of patients with lymphoma that achieved an accuracy of 0.87, while Jemaa et al. achieved a DSC of 0.873 in lymphoma and lung cancer dataset. While this was confirmed in our study using a mixed dataset of lung, lymphoma and melanoma patients, our observations also suggest a closer association between prediction accuracy and tumor to background ratio. From a clinical point of view the finding is quite interesting as it prospects the possibility of improving prediction models by adding SUV metrics as normal tissue radiotracer uptake or lesion-to-background ratio to analysis. While the model showed strong diagnostic performance with a high negative predictive value (NPV = 96%)—consistent with the expected predominance of true negative voxels in 3D PET imaging—its moderate positive predictive value (PPV = 83%) highlights the ongoing challenge of reducing false positives in metabolic lesion classification. These results indicate high confidence in ruling out disease (NPV = 96%) but suggest that 17% of PET-positive findings could be misclassified. The balanced accuracy of 80%, although comparable to current results from the AutoPET project [ 48 , 49 ], remains below the optimal threshold recommended for clinical use, especially for equivocal lesions (TBR 1.5-2.0) where higher sensitivity is critical to avoid false negatives. This study confirms that nnU-Net is a reliable approach for segmenting hypermetabolic lesions in [¹⁸F]FDG-PET/CT imaging, allowing an objective assessment of the global disease burden. We observed a high correlation between the predicted and manually annotated lesion volumes. These results are in line with those reported in previous papers [ 26 , 48 , 49 , 50 ], on a larger and more heterogeneous population composed of two datasets. Furthermore, our study aimed to evaluate the global and regional association between predictive performance and SUV uptake since segmentation of FDG-low uptake tumors remains difficult and associated with poor predictive performance [ 21 , 48 , 51 ]. Indeed, we observed a significant decrease (13%-21%) in prediction performance (Dice score) for low-uptake lesions (TBR ≤ 2) compared to high-uptake lesions. Moreover, regional analysis confirmed a significant association between TBR and segmentation accuracy as measured by DSCs. Conclusions This observation suggests the potential for evaluating automatic segmentation models that also include AI-based image enhancement tools to improve contrast and provide sharper images for segmentation. Furthermore, deep learning could be valuable for estimating the global burden of disease, offering a tool to support the assessment of therapeutic response and patient management. Declarations Acknowledgements We thank the staff of the PET group, the information systems unit and the General Management of the S. Andrea Hospital for the assistance provided during this research work. Author contributions AC conceived the study, participated in its design, implemented script under python/R packages, drafted the manuscript and made the final revision. NY, and DM performed image quality check and participating to critical review of data analysis and approved final version of manuscript. BA performed statistical analysis, OF, and FF verified and prepared imaging and clinical/demographic databases for further analysis. AM partecipated to critical review of data analysis and approved final version of manuscript. MCn, GG, participated in the study design, participating to critical review of data analysis. LM, LSM, and MCs performed image manual annotation, participating to critical review of final draft. Funding None Availability of data and material autoPET dataset analysed in this study is part of the MICCAI autoPET challenge 2022 https://autopet.grand-challenge.org/. The LOMICS20 dataset is not publicly available due to patient clinical data protection concerns. Ethical approval and Consent to participate Study was approved by the institutional ethics committee of the Medical Faculty of the University of Tübingen as well as the institutional data security and privacy review board and regional ethics committees Study was conducted in accordance with the first revision of the Declaration of Helsinki from 1975. Informed consent was obtained from all individual participants included in the study. Consent for publication Not applicable. Competing Interests The authors declare that they have no competing interests. References Seymour L, Bogaerts J, Perrone A, Ford R, Schwartz LH, Mandrekar S, et al. iRECIST: guidelines for response criteria for use in trials testing immunotherapeutics. The Lancet Oncology. 2017;18:e143-e52. doi:10.1016/S1470-2045(17)30074-8. Persigehl T, Lennartz S, Schwartz LH. iRECIST: how to do it. Cancer imaging : the official publication of the International Cancer Imaging Society. 2020;20:2. doi:10.1186/s40644-019-0281-x. Bleker J, Kwee TC, Rouw D, Roest C, Borstlap J, de Jong IJ, et al. A deep learning masked segmentation alternative to manual segmentation in biparametric MRI prostate cancer radiomics. European radiology. 2022;32:6526-35. doi:10.1007/s00330-022-08712-8. Paez D, Giammarile F, Orellana P. Nuclear medicine: a global perspective. Clinical and Translational Imaging. 2020;8:51-3. doi:10.1007/s40336-020-00359-z. OECD. Health care utilization–diagnostic exams,. https://statsoecdorg/indexaspx?queryid=30160. 2022. Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA: a cancer journal for clinicians. 2021;71:209-49. doi:10.3322/caac.21660. Safiri S, Kolahi AA, Naghavi M, Global Burden of Disease Bladder Cancer C. Global, regional and national burden of bladder cancer and its attributable risk factors in 204 countries and territories, 1990-2019: a systematic analysis for the Global Burden of Disease study 2019. BMJ global health. 2021;6. doi:10.1136/bmjgh-2020-004128. Ebner R, Sheikh GT, Brendel M, Ricke J, Cyran CC. ESR Essentials: staging and restaging with FDG-PET/CT in oncology-practice recommendations by the European Society for Hybrid, Molecular and Translational Imaging. European radiology. 2024. doi:10.1007/s00330-024-11094-8. Ebner R, Sheikh GT, Brendel M, Ricke J, Cyran CC. ESR Essentials: role of PET/CT in neuroendocrine tumors-practice recommendations by the European Society for Hybrid, Molecular and Translational Imaging. European radiology. 2025;35:1903-12. doi:10.1007/s00330-024-11095-7. Alotaibi NA, Yakar D, Glaudemans A, Kwee TC. Diagnostic errors in clinical FDG-PET/CT. Eur J Radiol. 2020;132:109296. doi:10.1016/j.ejrad.2020.109296. Toxopeus R, Kasalak Ö, Yakar D, Noordzij W, Dierckx R, Kwee TC. Is work overload associated with diagnostic errors on (18)F-FDG-PET/CT? European journal of nuclear medicine and molecular imaging.2024 Mar;51(4):1079-1084. doi:10.1007/s00259-023-6543-3. Yang F, Zamzmi G, Angara S, Rajaraman S, Aquilina A, Xue Z, et al. Assessing Inter-Annotator Agreement for Medical Image Segmentation. IEEE access : practical innovations, open solutions. 2023;11:21300-12. doi:10.1109/access.2023.3249759. Tran KA, Kondrashova O, Bradley A, Williams ED, Pearson JV, Waddell N. Deep learning in cancer diagnosis, prognosis and treatment selection. Genome medicine. 2021;13:152. doi:10.1186/s13073-021-00968-x. Liao J, Li X, Gan Y, Han S, Rong P, Wang W, et al. Artificial intelligence assists precision medicine in cancer treatment. Frontiers in oncology. 2022;12:998222. doi:10.3389/fonc.2022.998222. Ciarmiello A, Giovannini E, Tutino F, Yosifov N, Milano A, Florimonte L, et al. Does FDG PET-Based Radiomics Have an Added Value for Prediction of Overall Survival in Non-Small Cell Lung Cancer? Journal of clinical medicine. 2024;13. doi:10.3390/jcm13092613. Shiyam Sundar LK, Yu J, Muzik O, Kulterer OC, Fueger B, Kifjak D, et al. Fully Automated, Semantic Segmentation of Whole-Body (18)F-FDG PET/CT Images Based on Data-Centric Artificial Intelligence. Journal of nuclear medicine : official publication, Society of Nuclear Medicine. 2022;63:1941-8. doi:jnumed.122.264063 [pii]264063 [pii]10.2967/jnumed.122.264063. Wasserthal J, Breit HC, Meyer MT, Pradella M, Hinck D, Sauter AW, et al. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiology Artificial intelligence. 2023;5:e230024. doi:10.1148/ryai.230024. Hasani N, Paravastu SS, Farhadi F, Yousefirizi F, Morris MA, Rahmim A, et al. Artificial Intelligence in Lymphoma PET Imaging:: A Scoping Review (Current Trends and Future Directions). PET Clin. 2022;17:145-74. doi: 10.1016/j.cpet.2021.09.006. Häggström I, Leithner D, Alvén J, Campanella G, Abusamra M, Zhang H, et al. Deep learning for [18F]fluorodeoxyglucose-PET-CT classification in patients with lymphoma: a dual-centre retrospective analysis. The Lancet Digital Health. 2024;6:e114-e25. doi:10.1016/s2589-7500(23)00203-0. Gatidis S, Früh M, Fabritius MP, Gu S, Nikolaou K, Fougère CL, et al. Results from the autoPET challenge on fully automated lesion segmentation in oncologic PET/CT imaging. Nature Machine Intelligence. 2024;6:1396-405. doi:10.1038/s42256-024-00912-9. Tarai S, Lundström E, Ahmad N, Strand R, Ahlström H, Kullberg J. Whole-body tumor segmentation from FDG-PET/CT: Leveraging a segmentation prior from tissue-wise projections. Heliyon. 2025;11:e41038. doi:https://doi.org/10.1016/j.heliyon.2024.e41038. Rahman WT, Wale DJ, Viglianti BL, Townsend DM, Manganaro MS, Gross MD, et al. The impact of infection and inflammation in oncologic (18)F-FDG PET/CT imaging. Biomedicine & pharmacotherapy = Biomedecine & pharmacotherapie. 2019;117:109168. doi:10.1016/j.biopha.2019.109168. Hess S, Scholtens AM, Gormsen LC. Patient Preparation and Patient-related Challenges with FDG-PET/CT in Infectious and Inflammatory Disease. PET Clin. 2020;15:125-34. doi:10.1016/j.cpet.2019.11.001. Metser U, Even-Sapir E. Increased (18)F-fluorodeoxyglucose uptake in benign, nonphysiologic lesions found on whole-body positron emission tomography/computed tomography (PET/CT): accumulated data from four years of experience with PET/CT. Semin Nucl Med. 2007;37:206-22. doi:10.1053/j.semnuclmed.2007.01.001. Gatidis S, T K. A whole-body FDG-PET/CT dataset with manually annotated tumor lesions (FDG-PET-CT-Lesions) [Dataset]. The Cancer Imaging Archive. 2022. doi: 10.7937/gkr0-xv29 Gatidis S, Hepp T, Fruh M, La Fougere C, Nikolaou K, Pfannenberg C, et al. A whole-body FDG-PET/CT Dataset with manually annotated Tumor Lesions. Scientific data. 2022;9:601. doi:10.1038/s41597-022-01718-3 [pii]1718 [pii]10.1038/s41597-022-01718-3. Boellaard R, Delgado-Bolton R, Oyen WJ, Giammarile F, Tatsch K, Eschner W, et al. FDG PET/CT: EANM procedure guidelines for tumour imaging: version 2.0. European journal of nuclear medicine and molecular imaging. 2015;42:328-54. doi:10.1007/s00259-014-2961-x. Ronneberger O. Medical image computing and computer-assisted intervention–MICCAI 2015. Springer; 2015. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods. 2021;18:203-11. doi:10.1038/s41592-020-01008-z. Yu C, Anakwenze CP, Zhao Y, Martin RM, Ludmir EB, J SN, et al. Multi-organ segmentation of abdominal structures from non-contrast and contrast enhanced CT images. Scientific reports. 2022;12:19093. doi:10.1038/s41598-022-21206-3 [pii]21206 [pii]10.1038/s41598-022-21206-3. Xue H, Fang Q, Yao Y, Teng Y. 3D PET/CT tumor segmentation based on nnU-Net with GCN refinement. Phys Med Biol. 2023;68. doi:10.1088/1361-6560/acede6. Kalisch H, Hörst F, K H, Kleesiek J, Seibold C. Autopet III challenge: Incorporating anatomical knowledge into nnUNet for lesion segmentation in PET/CT. arXiv preprint arXiv:221102701. 2024;2024. Astaraki M, Bendazzolie S. Lesion Segmentation in Whole-Body Multi-Tracer PET-CT Images; a Contribution to AutoPET 2024 Challenge. arXiv preprint arXiv:221102701. 2024. Virtanen P, Gommers R, Oliphant TE, Haberland M, Reddy T, Cournapeau D, et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nature methods. 2020;17:261-72. doi:10.1038/s41592-019-0686-2. Cardoso MJ, Li W, Brown R, Ma N, Kerfoot E, Wang Y, et al. Monai: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:221102701. 2022. Dice LR. Measures of the amount of ecologic association between species. Ecology. 1945;26:297-302. Sachpekidis C, Enqvist O, Ulen J, Kopp-Schneider A, Pan L, Jauch A, et al. Application of an artificial intelligence-based tool in [(18)F]FDG PET/CT for the assessment of bone marrow involvement in multiple myeloma. European journal of nuclear medicine and molecular imaging. 2023;50:3697-708. doi:10.1007/s00259-023-06339-5. Yousefirizi F, Klyuzhin IS, O JH, Harsini S, Tie X, Shiri I, et al. TMTV-Net: fully automated total metabolic tumor volume segmentation in lymphoma PET/CT images - a multi-center generalizability analysis. Eur J Nucl Med Mol Imaging. 2024;51:1937-54. doi: 10.007/s00259-024-6616-x. Epub 2024 Feb 8. Jemaa S, Fredrickson J, Carano RAD, Nielsen T, de Crespigny A, Bengtsson T. Tumor Segmentation and Feature Extraction from Whole-Body FDG-PET/CT Using Cascaded 2D and 3D Convolutional Neural Networks. Journal of digital imaging. 2020;33:888-94. doi:10.1007/s10278-020-00341-1. Huang B, Chen Z, Wu PM, Ye Y, Feng ST, Wong CO, et al. Fully Automated Delineation of Gross Tumor Volume for Head and Neck Cancer on PET-CT Using Deep Learning: A Dual-Center Study. Contrast media & molecular imaging. 2018;2018:8923028. doi:10.1155/2018/8923028. Teramoto A, Fujita H, Yamamuro O, Tamaki T. Automated detection of pulmonary nodules in PET/CT images: Ensemble false-positive reduction using a convolutional neural network technique. Medical physics. 2016;43:2821-7. doi:10.1118/1.4948498. Zincirkeser S, Sahin E, Halac M, Sager S. Standardized uptake values of normal organs on 18F-fluorodeoxyglucose positron emission tomography and computed tomography imaging. J Int Med Res. 2007;35:231-6. doi:10.1177/147323000703500207. Sharma P, Chatterjee P, Alvarado L, Dwivedi A. Standardized uptake value of normal organs on routine clinical [18F]FDG PET/CT: impact of tumor metabolism and patient-related factors. Nuclear Medicine Review. 2023;26:1-10. doi:10.5603/NMR.a2022.0036. Lindgren Belal S, Sadik M, Kaboteh R, Enqvist O, Ulén J, Poulsen MH, et al. Deep learning for segmentation of 49 selected bones in CT scans: First step in automated PET/CT-based 3D quantification of skeletal metastases. Eur J Radiol. 2019;113:89-95. Carles M, Kuhn D, Fechter T, Baltas D, Mix M, Nestle U, et al. Development and evaluation of two open-source nnU-Net models for automatic segmentation of lung tumors on PET and CT images with and without respiratory motion compensation. Eur Radiol. 2024;34:6701-11. doi: 10.1007/s00330-024-10751-2. Epub 2024 Apr 25. Jafari E, Zarei A, Dadgar H, Keshavarz A, Manafi-Farid R, Rostami H, et al. A convolutional neural network-based system for fully automatic segmentation of whole-body [(68)Ga]Ga-PSMA PET images in prostate cancer. Eur J Nucl Med Mol Imaging. 2024;51:1476-87. doi: 10.007/s00259-023-6555-z. Epub 2023 Dec 14. Im HJ, Bradshaw T, Solaiyappan M, Cho SY. Current Methods to Define Metabolic Tumor Volume in Positron Emission Tomography: Which One is Better? Nucl Med Mol Imaging. 2018;52:5-15. doi: 0.1007/s13139-017-0493-6. Epub 2017 Sep 19. Andrearczyk V, Oreiller V, Boughdad S, Le Rest CC, Tankyevych O, Elhalawani H, et al. Automatic Head and Neck Tumor segmentation and outcome prediction relying on FDG-PET/CT images: Findings from the second edition of the HECKTOR challenge. Med Image Anal. 2023;90:102972. doi:S1361-8415(23)00232-3 [pii]10.1016/j.media.2023.102972. Oreiller V, Andrearczyk V, Jreige M, Boughdad S, Elhalawani H, Castelli J, et al. Head and neck tumor segmentation in PET/CT: The HECKTOR challenge. Med Image Anal. 2022;77:102336. doi:S1361-8415(21)00381-9 [pii]10.1016/j.media.2021.102336. Hatt M, Laurent B, Ouahabi A, Fayad H, Tan S, Li L, et al. The first MICCAI challenge on PET tumor segmentation. Med Image Anal. 2018;44:177-95. doi:S1361-8415(17)30189-5 [pii]10.1016/j.media.2017.12.007. Moreau N, Rousseau C, Fourcade C, Santini G, Ferrer L, Lacombe M, et al. Influence of inputs for bone lesion segmentation in longitudinal (18)F-FDG PET/CT imaging studies. Annual International Conference of the IEEE Engineering in Medicine and Biology Society IEEE Engineering in Medicine and Biology Society Annual International Conference. 2022;2022:4736-9. doi:10.1109/EMBC48229.2022.9871081. Tables Tables 1 to 3 are available in the Supplementary Files section. Supplementary Files FigureS1.png Supplemental data Figure 1. Dice loss and score function values per epoch. The figure shows the dice loss and dice score curves obtained by the nnU-Net segmentation model on the training and validation datasets, respectively. The curves show a decrease in dice loss and an increase in dice score as the number of epochs increases. table1.pdf table2.pdf table3.pdf Cite Share Download PDF Status: Published Journal Publication published 28 Nov, 2025 Read the published version in EJNMMI Research → Version 1 posted Editorial decision: Major Revision 25 Aug, 2025 Reviewers agreed at journal 18 Jul, 2025 Reviewers invited by journal 13 Jul, 2025 Editor assigned by journal 26 Jun, 2025 First submitted to journal 21 Jun, 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-6895938","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":484745515,"identity":"63473399-8f72-4c97-ab1d-095bbd18b816","order_by":0,"name":"Andrea Ciarmiello","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA10lEQVRIiWNgGAWjYDACZhBhwCDDx8B8gIGxgQQtPGwMbAlEaoECoBYeA+K08LczP/t0o8CGh42955vEzx02cgzsh49uwKdF4jCb8ewcgzQeNp6z2yR7z6QZM/Ckpd3Ap8WAmcGYOcfgMA+bRO42Cd62w4kNEjxmBLSwfwZq+Q/UkvNM8i9xWnhAthwAaWGTJsoWicM8xUAtyUC/HDO2lm1LM2Yj5Bf+/uObmXP+2Mnxszc/vPm2zQbIOHwMrxZkwCIBItmIVQ4CzB9IUT0KRsEoGAUjBwAAuxg9B9m1UEoAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0002-6579-656X","institution":"Nuclear Medicine Department, S. Andrea Hospital, La Spezia","correspondingAuthor":true,"prefix":"","firstName":"Andrea","middleName":"","lastName":"Ciarmiello","suffix":""},{"id":484745516,"identity":"48aa3b98-130a-42c0-a294-397436d6b022","order_by":1,"name":"Nikola Yosifov","email":"","orcid":"","institution":"Nuclear Medicine Department, S.Andrea Hospital, La Spezia","correspondingAuthor":false,"prefix":"","firstName":"Nikola","middleName":"","lastName":"Yosifov","suffix":""},{"id":484745517,"identity":"c9f428fb-4ab7-4d5f-893c-01ca83eff275","order_by":2,"name":"Donatella Masciale","email":"","orcid":"","institution":"Nuclear Medicine Department, S. Andrea Hospital, La Spezia","correspondingAuthor":false,"prefix":"","firstName":"Donatella","middleName":"","lastName":"Masciale","suffix":""},{"id":484745518,"identity":"232f623b-416d-493e-8c1a-a0a7e34f48a1","order_by":3,"name":"Ornella Ferrando","email":"","orcid":"","institution":"Medical Physics Department, S. Andrea Hospital , La Spezia","correspondingAuthor":false,"prefix":"","firstName":"Ornella","middleName":"","lastName":"Ferrando","suffix":""},{"id":484745519,"identity":"eb71f77e-93a0-4d65-8bb0-16f6393b4ec8","order_by":4,"name":"Franca Foppiano","email":"","orcid":"","institution":"Medical Physics Department, S. Andrea Hospital, La Spezia","correspondingAuthor":false,"prefix":"","firstName":"Franca","middleName":"","lastName":"Foppiano","suffix":""},{"id":484745520,"identity":"4444f925-9edb-4bc4-88c5-f2c448a673dd","order_by":5,"name":"Amalia Milano","email":"","orcid":"","institution":"Oncology Uint, S. Andrea Hospital, La Spezia","correspondingAuthor":false,"prefix":"","firstName":"Amalia","middleName":"","lastName":"Milano","suffix":""},{"id":484745521,"identity":"10047bc5-3a6b-4c30-80bc-e6d5c2b01f48","order_by":6,"name":"Massimo Canevari","email":"","orcid":"","institution":"Biomedical Engineering Department, S. Andrea Hospital, La Spezia","correspondingAuthor":false,"prefix":"","firstName":"Massimo","middleName":"","lastName":"Canevari","suffix":""},{"id":484745522,"identity":"e4e47baa-a8ad-48a2-8405-1aa8de537c64","order_by":7,"name":"Luigia Florimonte","email":"","orcid":"","institution":"Nuclear Medicine Unit, Fondazione IRCCS Ca' Granda Ospedale Maggiore Policlinico","correspondingAuthor":false,"prefix":"","firstName":"Luigia","middleName":"","lastName":"Florimonte","suffix":""},{"id":484745523,"identity":"14da1603-1b62-49cc-bde0-ba6529eb40f1","order_by":8,"name":"Massimo Castellani","email":"","orcid":"","institution":"Nuclear Medicine Unit, Fondazione IRCCS Ca' Granda Ospedale Maggiore Policlinico","correspondingAuthor":false,"prefix":"","firstName":"Massimo","middleName":"","lastName":"Castellani","suffix":""},{"id":484745524,"identity":"38561fd4-f9c3-45b0-974b-8e8b939a5335","order_by":9,"name":"Giampiero Giovacchini","email":"","orcid":"","institution":"Nuclear Medicine Department. S. Andrea Hospital. La Spezia","correspondingAuthor":false,"prefix":"","firstName":"Giampiero","middleName":"","lastName":"Giovacchini","suffix":""},{"id":484745525,"identity":"2b7ef291-28eb-47bb-9577-4a446f171f59","order_by":10,"name":"Lorenzo Stefano Maffioli","email":"","orcid":"","institution":"IRCCS Istituto Romagnolo per lo Studio dei Tumori Dino Amadori","correspondingAuthor":false,"prefix":"","firstName":"Lorenzo","middleName":"Stefano","lastName":"Maffioli","suffix":""},{"id":484745526,"identity":"7ccebbb0-5e8c-4c1d-93c4-7f3accc7aaef","order_by":11,"name":"Bruno Alfano","email":"","orcid":"","institution":"Human Shape Technologies","correspondingAuthor":false,"prefix":"","firstName":"Bruno","middleName":"","lastName":"Alfano","suffix":""}],"badges":[],"createdAt":"2025-06-15 00:50:18","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6895938/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6895938/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s13550-025-01333-4","type":"published","date":"2025-11-28T15:57:20+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":87033180,"identity":"152d04ea-c624-4ed3-bfac-4cfc9661a39b","added_by":"auto","created_at":"2025-07-18 13:04:03","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":584907,"visible":true,"origin":"","legend":"\u003cp\u003eQuantitative assessment of deep learning-based segmentation of hypermetabolic lesions from [¹⁸F]FDG-PET/CT images.\u003c/p\u003e\n\u003cp\u003ePanel A: Probability density (y-axis) of the Dice score (x-axis) obtained from unsupervised segmentation. No significant differences were found between the estimated probability density in the validation and test datasets. Panel B: Relationship between predicted tumor volumes and manually labeled tumor volumes in positive scans.\u003c/p\u003e","description":"","filename":"figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-6895938/v1/0a356b2ef69b93efcc5a863c.png"},{"id":87035046,"identity":"0c383e2c-217e-4a75-8a1d-4bdb20aad5c6","added_by":"auto","created_at":"2025-07-18 13:12:03","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":68535,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative image of automatic segmentation of hypermetabolic lesions.\u003c/p\u003e\n\u003cp\u003ePanel A: [¹⁸F]FDG-PET/CT images showing an area of ​​intense uptake in the upper lobe of the right lung. Panel B: The grayscale background shows the anatomical structures segmented by the CT component of the PET-CT scan. The color overlay image shows the automatic segmentation map. True and false positives are shown in red and cyan, respectively.\u003c/p\u003e","description":"","filename":"figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-6895938/v1/b09a6aaa8696bb8b0b03ca23.png"},{"id":87035047,"identity":"c3f20a15-f6e8-404d-8445-087e015b9f13","added_by":"auto","created_at":"2025-07-18 13:12:03","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":603319,"visible":true,"origin":"","legend":"\u003cp\u003ePrediction performance by anatomical region.\u003c/p\u003e\n\u003cp\u003ePanel A: Histogram displaying the number of true positives and false negatives by anatomical region.\u003c/p\u003e\n\u003cp\u003ePanel B shows the correlation between dice score and the tumor to background ratio. Marker size indicates the number of lesions detected at each anatomical site.\u003c/p\u003e","description":"","filename":"figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-6895938/v1/1d0132c92e9c3ad7d0f19e16.png"},{"id":97179112,"identity":"58f8c04e-d3c2-4eea-b857-0c021718d823","added_by":"auto","created_at":"2025-12-01 16:14:12","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1696615,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6895938/v1/7ce80ed3-b417-48a4-9250-3f09ce2e290a.pdf"},{"id":87033184,"identity":"e52898c7-4931-4537-8be9-e2dc342f2712","added_by":"auto","created_at":"2025-07-18 13:04:03","extension":"png","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":525364,"visible":true,"origin":"","legend":"\u003cp\u003eSupplemental data\u003c/p\u003e\n\u003cp\u003eFigure 1. Dice loss and score function values per epoch.\u003c/p\u003e\n\u003cp\u003eThe figure shows the dice loss and dice score curves obtained by the nnU-Net segmentation model on the training and validation datasets, respectively. The curves show a decrease in dice loss and an increase in dice score as the number of epochs increases.\u003c/p\u003e","description":"","filename":"FigureS1.png","url":"https://assets-eu.researchsquare.com/files/rs-6895938/v1/bc37dcd24f2d2e32d054a385.png"},{"id":87033178,"identity":"c7be811f-d75e-40fe-9c50-319de695381e","added_by":"auto","created_at":"2025-07-18 13:04:03","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":63158,"visible":true,"origin":"","legend":"","description":"","filename":"table1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6895938/v1/34008f0a940831f20a777c54.pdf"},{"id":87033190,"identity":"7ba7e415-e41e-4ccf-823d-1b00348f939e","added_by":"auto","created_at":"2025-07-18 13:04:03","extension":"pdf","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":48384,"visible":true,"origin":"","legend":"","description":"","filename":"table2.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6895938/v1/56b3a330fdb18de05ff98416.pdf"},{"id":87033182,"identity":"b8c6dbdb-0db5-4c2f-8e0e-f9784d357d18","added_by":"auto","created_at":"2025-07-18 13:04:03","extension":"pdf","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":46827,"visible":true,"origin":"","legend":"","description":"","filename":"table3.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6895938/v1/d239aaefbdc1cb629e2b1430.pdf"}],"financialInterests":"","formattedTitle":"\u003cp\u003eGlobal and regional accuracy of deep learning-based tumor segmentation from whole-body [¹⁸F]fluorodeoxyglucose PET/CT images\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eStaging of tumor lesions and the monitoring of response to many cancer therapy regimens including immune checkpoint inhibitors is based on [\u0026sup1;⁸F]fluorodeoxyglucose ([\u0026sup1;⁸F]FDG)-PET/CT images signaling areas of high glucose uptake[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The lesions are accurately measured to establish the patient\u0026rsquo;s overall disease burden, by means of manual segmentation performed by a highly trained specialists (radiologist or nuclear medicine physician). The process is painstaking and may require up to 5 minutes for a single lesion and several hours in patients with multiple lesions [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], thus occupying a large portion of the workload of nuclear medicine physicians.\u003c/p\u003e\u003cp\u003eWith the increasing trend of oncological malignancies and increase in the number of PET/CT scans estimated to be the 13% in the last decade alone [\u003cspan additionalcitationids=\"CR5 CR6\" citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] several efforts have been made to alleviate this burden with the development in advanced imaging technology and non-invasive diagnostic techniques [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eNonetheless, many tasks involved in staging and follow-up remain operator-dependent and as such are subject to errors [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], and inter- and intra-observer variability [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Recently, the introduction of several deep learning algorithms have provided additional tools for quicker and more accurate tumor assessment to support patient-tailored treatments [\u003cspan additionalcitationids=\"CR14\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eArtificial intelligence (AI) with deep neural network (DNN) frameworks has already been used for segmentation of normal anatomical structures using computed tomography (CT) images [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. To date, the most convincing open-source software is that offered by Wasserthal et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] where a DNN model is able to segment 117 anatomical structures.\u003c/p\u003e\u003cp\u003eAI-based DNNs models have also been proposed to automatically segment hypermetabolic malignant lesions from [\u0026sup1;⁸F]FDG-PET/CT images [\u003cspan additionalcitationids=\"CR19 CR20\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] or to correctly classify hypermetabolic regions of non-tumor nature as in the case of physiologically hypermetabolic tissues or high uptake due to infectious and inflammatory diseases [\u003cspan additionalcitationids=\"CR23\" citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThis task is challenging as in the first case, the high metabolic activity of the healthy background tissue could limit the model's ability to recognize the tumor lesion, while in the latter case the model could classify a benign pathological process as malignant.\u003c/p\u003e\u003cp\u003eIn order to evaluate the global and anatomical region-specific performance of the convolutional neural network segmentation model, we retrospectively studied its diagnostic accuracy using a sample of 3D [\u0026sup1;⁸F]FDG-PET/CT scans of patients diagnosed with lymphoma, melanoma, and lung cancer. The primary objective of our study was to test the relationship between the neural network predicted and manually segmented volumes. The secondary objective was to measure the extent at which the predictive accuracy is associated with normal background uptake.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eDataset\u003c/h2\u003e\u003cp\u003eThis retrospective study was carried on FDG-PET/CT images from the autoPET dataset [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] and the LOMICS20 dataset [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Both studies have been approved by the respective regional ethics committees.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eautoPET dataset\u003c/h3\u003e\n\u003cp\u003eThis dataset includes FDG-PET/CT scans of 145 lymphomas, 168 lung cancers, and 188 melanomas, as well as a 513 scans of negative controls. Negative PET/CT scans came from patients studied after medical or surgical treatment with no evidence of metabolically active disease. Image voxel size is (2.04 x 2.04 x 3.00) mm\u0026sup3;. 3D volumes of manual annotations are provided for all PET/CT scan. Manual annotation was made by two experienced radiologists with ten and five years of experience. autoPet cohort is publicly available at TCIA (The Cancer Imaging Archive) [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003eLOMICS20 dataset\u003c/h3\u003e\n\u003cp\u003eA subset of images from the LOMICS20 cohort consisting of 145 whole-body FDG-PET/CT scans of lung cancer patients. Image voxel size is (2.73 x 2.73 x 3.27) mm\u0026sup3;. 3D volumes of manual annotations provided for all PET/CT scan was performed by two board-certified nuclear medicine physicians with five years of experience. LOMICS20 Dataset is not publicly available.\u003c/p\u003e\u003cp\u003eIn order to train, validate and test the prediction model on different samples, the cohort was randomly split into three datasets. Therefore, a total of 881 [\u0026sup1;⁸F]FDG-PET/CT scans (70% of the entire dataset) were used as the training set, and 232 scans (20% of the dataset) were used for internal validation. Additionally, a subsample of 116 scans (10% of the dataset), which were not part of the training or validation sets, was designated as the test dataset (Table\u0026nbsp;1). Random splitting was performed with the sklearn platform in Python 3 to obtain datasets comparable by age, gender, diagnosis and cases with hypermetabolic lesions.\u003c/p\u003e\n\u003ch3\u003eProcedures\u003c/h3\u003e\n\u003cp\u003ePET/CT acquisition and processing was performed in agreement with international guidelines for PET/CT examinations in oncology [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. The radiotracer [\u0026sup1;⁸F]FDG was administered after 6 hours of fasting and blood glucose level check in all patients. Scans were performed at least 60 minutes after its administration. Details on acquisition and image reconstruction parameters are reported elsewhere [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eAs by clinical standards, PET/CT images were classified as positive in the presence of one or more lesions featuring high [\u0026sup1;⁸F]FDG uptake. The positive sample consisted of 188 melanomas, 313 lung tumors, and 145 lymphomas.\u003c/p\u003e\u003cp\u003eThe PET/CT images were reviewed by board-certified nuclear medicine physicians who manually segmented positive lesions using dedicated software, NORA image analysis platform (University of Freiburg, Germany), or PET Volume Computerized Assisted Reporting (PETVCAR) \u0026ndash;a commercial software running on Advantage Workstation (version 4.6; GE Healthcare). The resulting 3D binary mask was used as ground-truth in the training, validation and testing phases.\u003c/p\u003e\n\u003ch3\u003eNetwork model\u003c/h3\u003e\n\u003cp\u003eThe segmentation model was implemented in the nnU-Net framework. This is a U-Net based network able of configuring all hyperparameters on the dataset feature [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. This architecture was selected for its reported high performance in segmenting both normal and pathological 3D medical images from different imaging modalities [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]\u003c/p\u003e\u003cp\u003eThe network architecture consisting of six stages in the encoder, with (1, 3, 4, 6, 6, 6) blocks in each stage, and (16, 32, 64, 128, 256, 320) feature layers per stage, respectively and the Leaky ReLU activation function.The network parameters were optimized over 1000 epochs using the Adam optimizer, with a learning rate of 1e-4, a weight decay of 1e-5, a dropout rate of 0.20, and a batch size of 2. Patches of (96\u0026times;128\u0026times;128) voxels were extracted through a sliding window approach with an overlap of 0.25 between consecutive patches [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. To avoid overfitting, an early stopping strategy with a patience of 60 epochs and a delta of 0.01 is implemented.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003ePreprocessing and data augumentation\u003c/h2\u003e\u003cp\u003eCT images were downsampled to match the corresponding PET image size. PET, CT, and label images were then resampled to a common imaging resolution, with uniform isotropic spacing of 2 mm\u0026sup3; using bilinear interpolation for PET and CT images and nearest-neighbor interpolation for label images. PET images were then processed by applying Min-Max normalization to standardized uptake values ​​(SUV) [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. To augment the input to the network we applied at start of every epoch a series of random transformations to PET and label images including a) 3D elastic deformations applying a Gaussian kernel with standard deviation and offsets uniformly sampled from (0, 1), b) random scaling by 1.1 factor along axes, c) axial rotations by (\u0026minus;π/12, π/12) angle, d) 3D translations in the range (0, 10) voxels along axes, e) addition of random Gaussian noise (\u0026micro;\u0026thinsp;=\u0026thinsp;0 ,σ\u0026thinsp;=\u0026thinsp;1) and f) Gamma correction (γ \u0026isin; 0.7, 1.5) [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eLesions anatomical mapping\u003c/h3\u003e\n\u003cp\u003eTotalSegmentator was used to segment anatomical structures from PET/CT images to localize hypermetabolic lesions and assess tumor-to-background ratio (TBR) for quantitative analysis (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/wasserth/TotalSegmentator\u003c/span\u003e\u003cspan address=\"https://github.com/wasserth/TotalSegmentator\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eLesions were mapped onto the corresponding anatomical structure by matching the lesion center of mass with the spatial coordinates of the anatomical map.\u003c/p\u003e\u003cp\u003eLesions whose center of mass did not match the coordinates of any extracted tissue were used for global assessment only and not included in the regional analysis. Lesions mapping was performed with Scipy ndimage library (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/scipy/scipy\u003c/span\u003e\u003cspan address=\"https://github.com/scipy/scipy\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003eSUV estimate\u003c/h3\u003e\n\u003cp\u003eSUV images were used to calculate metabolic tumor volume (MTV) and total lesion glycolysis (TLG). Tumor-to-background ratio (TBR) was calculated by dividing the lesion SUVmax by the SUVmean of corresponding background tissue. Background activity was measured on the lesion's anatomical structure after masking the region corresponding to the hypermetabolic area as defined by annotated label.\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eDevelopment environment\u003c/h2\u003e\u003cp\u003eThe project was carried out using the open-source platform MONAI [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e] (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps:/monai.io\u003c/span\u003e\u003cspan address=\"https://monai.io\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), which is designed to standardize medical image processing by means of a broad selection of AI-based processing tools including image pre-processing, training, and post-processing operations needed in the formulation of prediction models.\u003c/p\u003e\u003cp\u003eImage preprocessing and network implementation were performed in Python version 3.11.5 and cuda 12.2, using Python-based image processing libraries, such as Scikit-image version 0.20.0 (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://scikit-image.org/\u003c/span\u003e\u003cspan address=\"https://scikit-image.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), OpenCV version 4.6.0 (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://opencv.org/\u003c/span\u003e\u003cspan address=\"https://opencv.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), and SciPy version 1.11.3 [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://scipy.org/\u003c/span\u003e\u003cspan address=\"https://scipy.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e).\u003c/p\u003e\u003cp\u003ePython scripts to convert images from DICOM to NIFTI and to generate CT volume resampled to PET, and standardized uptake values (SUV) images were available at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/lab-midas/TCIA\u003c/span\u003e\u003cspan address=\"https://github.com/lab-midas/TCIA\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e processing.\u003c/p\u003e\u003cp\u003eAll experiments were performed under Ubuntu 22.04.3 using a Windows 11 Pro 23H2 64-bit workstation with 128GB of RAM, a 2.40GHz, 24-core Intel Xeon Silver 4214R CPU. Processing-intensive tasks were handled by an NVIDIA RTX A4000 graphic card equipped with 16 GB RAM.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003eStatistics\u003c/h2\u003e\u003cp\u003eStatistical analysis was performed with RStudio 4.4.2 with the R libraries dplyr, tidyr, stringr, gtsummary, gt and tidyverse. Data were reported as median with range, unless otherwise specified. Categorical variables were reported as counts and percentages. Group comparisons used One-way analysis and Chi-square test for categorical variables. Correlations between manual and predicted segmentation were assessed using linear correlation. Spearman's rho was used to estimate the strength of association between prediction accuracy and TBR.\u003c/p\u003e\u003cp\u003eThe network was optimized using a Dice loss function [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e], with model performance quantified by the Dice Similarity Coefficient (DSC). During training, validation was performed at every epoch, with model selection based on maximal DSC achievement.\u003c/p\u003e\u003cp\u003eModel classification performance was evaluated using a confusion matrix (CM), comprising true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). From the CM Sensitivity (SS), Specificity (SP), Positive predictive value (PPV), Negative predictive value (NPV), and Balanced accuracy (BA) were derived. Performance was assessed across training, validation, and test datasets for PET/CT scans with and without hypermetabolic tumor sites. A p-value less than 0.05 was considered statistically significant.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eClinical and demographic characteristics of the study cohort are reported in Table\u0026nbsp;1. No differences were found in the distribution of demographic characteristics, diagnosis and positive cases number of the training, validation and test datasets (p\u0026thinsp;\u0026gt;\u0026thinsp;0.05).\u003c/p\u003e\u003cp\u003eAmong 1159 [\u0026sup1;⁸F]FDG-PET/CT scans in the study population, 646 (55.7%) showed positive findings for hypermetabolic lesions, with 7914 distinct high-glucose-metabolism lesions identified overall.\u003c/p\u003e\u003cp\u003eThe PET/CT segmentation model exhibited progressive convergence, with Dice loss decreasing consistently over 1000 training epochs (Supplementary Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). At completion, the model achieved a peak aggregate Dice similarity coefficient (DSC) of 0.805 on the validation set.\u003c/p\u003e\u003cp\u003eNo significant differences were found between the performance of the predictive model in the validation and test datasets (P\u0026thinsp;=\u0026thinsp;0.13). Density distribution of the dice score in the two groups is reported in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003e, panel A. In the [\u0026sup1;⁸F]FDG-positive cases, the manually annotated lesion volume significantly correlated with the predicted lesion volume (\u003cem\u003eR\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.88; p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001). Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003e, panel B shows the plot of the ground-truth volume versus the predicted volume.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows a representative image of automatic segmentation. Panel A shows [\u0026sup1;⁸F]FDG-PET/CT images with an area of ​​intense uptake in the upper lobe of the right lung. Panel B shows the grayscale map of segmented anatomical structures from the PET/CT images in the background. The overlaid color map shows true positive and false negative in red and cyan, respectively.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe proposed method achieved median Dice similarity coefficients of 0.70 (validation set) and 0.66 (test set) for lesion-level segmentation accuracy (Table\u0026nbsp;2).\u003c/p\u003e\u003cp\u003eIn the validation cohort, the model showed balanced accuracy of 0.82 (95% CI 0.74\u0026ndash;0.87), with positive predictive value (PPV) of 0.77 (95% CI 0.64\u0026ndash;0.85) and negative predictive value (NPV) of 0.97 (95% CI 0.94\u0026ndash;0.98).\u003c/p\u003e\u003cp\u003ePerformance in the test cohort showed sustained robustness, with a balanced accuracy of 0.79 (95% CI 0.69\u0026ndash;0.85), maintaining diagnostic reliability comparable to the validation set. The model achieved a PPV of 0.83 (95% CI 0.73\u0026ndash;0.86), and maintained NPV 0.96 (95% CI 0.93\u0026ndash;0.98) (Table\u0026nbsp;2).\u003c/p\u003e\u003cp\u003eOf 7914 hypermetabolic lesions, 5862 (78%) had a center of mass matching the spatial coordinates of key anatomical structures: adrenal glands, bones, brain, esophagus, kidneys, gallbladder, large and small intestines, liver, lungs, muscles, pancreas, prostate, spinal cord, spleen, stomach, thyroid, and trachea. Performance metrics, including true positive and false negative rates, showed significant inter-regional variability (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003e, panel A).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eA significant positive correlation was observed between the Dice score and tumor-to-background ratio (R\u0026thinsp;=\u0026thinsp;0.61, P\u0026thinsp;=\u0026thinsp;0.0082). Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003eB shows that the strength of this correlation varied by anatomical region, with higher Dice scores consistently associated with regions exhibiting elevated TBR. This relationship was further confirmed by one-way ANOVA, which showed significant differences in Dice scores across TBR-stratified groups (\u0026le;\u0026thinsp;1, 1\u0026ndash;2, \u0026gt;\u0026thinsp;2; p\u0026thinsp;\u0026lt;\u0026thinsp;0.001), as detailed in Table\u0026nbsp;3.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn recent years, several studies have confirmed the relevance of deep learning in the recognition and quantitative measurement of [\u0026sup1;⁸F]FDG-avid lesions in oncology and AI-based tools for automatized segmentation of hypermetabolic lesions from FDG-PET/CT scans [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eSome studies specifically including large populations of 5575 patients with lymphoma have reported lesion segmentation models able to achieve an accuracy of 0.875 [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Further studies have attempted to evaluate the accuracy of automatic segmentation on datasets including multiple oncological diseases such as Jemaa et al who reported a dice score of 0.873 on 3664 [\u0026sup1;⁸F]FDG-PET/CT scans [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eSo far, however, the majority of these results have been achieved from whole-body PET/CT scan datasets, where the overall lesion load is distributed over a significantly larger background, featuring significantly different radiotracer uptake across the anatomical structures included in the scan. These conditions might affect the predictive performance, resulting in differences in regional compared to global lesion accuracy.\u003c/p\u003e\u003cp\u003eSome studies have focused on oncological disease localized in specific anatomical regions using datasets consisting of 1 or 2 bed scans thus obtaining a smaller background region and presumably more homogeneous radiotracer uptake levels. Among these, Huang et al. achieved a dice score of 0.732 on a dataset of 22 head and neck cancer patients [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e] and Teramoto et al., obtained a dice score of 0.85 on a sample of 104 [\u0026sup1;⁸F]FDG-PET/CT of patients with solitary lung nodules [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]. Such studies, however, were based on small sample sizes, which is a potential drawback for deep learning applications. Furthermore, non-whole-body scans do not allow estimation of the global disease burden, which is one of the expected results of automatic segmentation.\u003c/p\u003e\u003cp\u003eAs previously reported, normal tissues can have significantly different levels of [\u0026sup1;⁸F]FDG-PET/CT uptake such as cerebrum, cerebellum, myocardium, tonsils, liver and spleen which generally have higher uptake than other tissues.\u003c/p\u003e\u003cp\u003eKnowledge of physiological uptake is therefore required for correct interpretation of PET/CT studies [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e, \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e]. Hence, our study also evaluated whether predictive performance was associated with physiological variations in normal background uptake across different tissue types.\u003c/p\u003e\u003cp\u003eBy means of the nnU-Net model implemented in the MONAI framework we were able to segment [\u0026sup1;⁸F]FDG-avid lesions from PET/CT scans. Model segmentation performance measured by the dice score achieved a DCSs of 0.805 and 0.784 on the validation and test datasets, respectively indicating a high overlap between the model segmented and the manually-annotated lesions.\u003c/p\u003e\u003cp\u003eThe correlation between lesion volume segmented by the model and manually-annotated lesion volumes was almost linear (R2\u0026thinsp;=\u0026thinsp;0.88, P\u0026thinsp;\u0026lt;\u0026thinsp;0.0001) and is in line with a previous study that has reported the linear correlation between lesion volumes detected by nnU-NET model and manual annotated volumes [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eMoreover, we found that TBR is significantly associated with prediction accuracy as measured by the dice score (TBR vs. dice score\u0026thinsp;=\u0026thinsp;0.61; P\u0026thinsp;=\u0026thinsp;0.0082), and that lower TBR values are associated with significant worse predictive performance.\u003c/p\u003e\u003cp\u003eAs to the AI-based assessment of global disease burden, this has been studied in previous work for several oncological diseases [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan additionalcitationids=\"CR45\" citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. Disease burden quantification implies the segmentation of all hypermetabolic lesions to measure metabolically active tumor volumes, mean and max lesion activity, total metabolic tumor volume, and total lesion glycolysis [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eFully-automated [\u0026sup1;⁸F]FDG-PET/CT image segmentation models are particularly useful for staging and assessment of treatment response in several malignancies where the measurement of the overall burden of disease is essential for patient management.\u003c/p\u003e\u003cp\u003eAmong these, H\u0026auml;ggstr\u0026ouml;m et al. implemented a deep learning-based model for PET/CT segmentation of patients with lymphoma that achieved an accuracy of 0.87, while Jemaa et al. achieved a DSC of 0.873 in lymphoma and lung cancer dataset.\u003c/p\u003e\u003cp\u003eWhile this was confirmed in our study using a mixed dataset of lung, lymphoma and melanoma patients, our observations also suggest a closer association between prediction accuracy and tumor to background ratio.\u003c/p\u003e\u003cp\u003eFrom a clinical point of view the finding is quite interesting as it prospects the possibility of improving prediction models by adding SUV metrics as normal tissue radiotracer uptake or lesion-to-background ratio to analysis.\u003c/p\u003e\u003cp\u003eWhile the model showed strong diagnostic performance with a high negative predictive value (NPV\u0026thinsp;=\u0026thinsp;96%)\u0026mdash;consistent with the expected predominance of true negative voxels in 3D PET imaging\u0026mdash;its moderate positive predictive value (PPV\u0026thinsp;=\u0026thinsp;83%) highlights the ongoing challenge of reducing false positives in metabolic lesion classification. These results indicate high confidence in ruling out disease (NPV\u0026thinsp;=\u0026thinsp;96%) but suggest that 17% of PET-positive findings could be misclassified.\u003c/p\u003e\u003cp\u003eThe balanced accuracy of 80%, although comparable to current results from the AutoPET project [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e], remains below the optimal threshold recommended for clinical use, especially for equivocal lesions (TBR 1.5-2.0) where higher sensitivity is critical to avoid false negatives.\u003c/p\u003e\u003cp\u003eThis study confirms that nnU-Net is a reliable approach for segmenting hypermetabolic lesions in [\u0026sup1;⁸F]FDG-PET/CT imaging, allowing an objective assessment of the global disease burden. We observed a high correlation between the predicted and manually annotated lesion volumes. These results are in line with those reported in previous papers [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e, \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e], on a larger and more heterogeneous population composed of two datasets.\u003c/p\u003e\u003cp\u003eFurthermore, our study aimed to evaluate the global and regional association between predictive performance and SUV uptake since segmentation of FDG-low uptake tumors remains difficult and associated with poor predictive performance [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e, \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e]. Indeed, we observed a significant decrease (13%-21%) in prediction performance (Dice score) for low-uptake lesions (TBR\u0026thinsp;\u0026le;\u0026thinsp;2) compared to high-uptake lesions. Moreover, regional analysis confirmed a significant association between TBR and segmentation accuracy as measured by DSCs.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThis observation suggests the potential for evaluating automatic segmentation models that also include AI-based image enhancement tools to improve contrast and provide sharper images for segmentation. Furthermore, deep learning could be valuable for estimating the global burden of disease, offering a tool to support the assessment of therapeutic response and patient management.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank the staff of the PET group, the information systems unit and the General Management of the S. Andrea Hospital for the assistance provided during this research work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAC conceived the study, participated in its design, implemented script under python/R packages, drafted the manuscript and made the final revision.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNY, and DM performed image quality check and participating to critical review of data analysis and approved final version of manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eBA performed statistical analysis, \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOF, and FF verified and prepared imaging and clinical/demographic databases for further analysis.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAM partecipated to critical review of data analysis and approved final version of manuscript.\u003c/p\u003e\n\u003cp\u003eMCn, GG, participated in the study design, participating to critical review of data analysis.\u003c/p\u003e\n\u003cp\u003eLM, LSM, and MCs performed image manual annotation, participating to critical review of final draft.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNone\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and material\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eautoPET dataset analysed in this study is part of the MICCAI autoPET challenge 2022 https://autopet.grand-challenge.org/.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe LOMICS20 dataset is not publicly available due to patient clinical data protection concerns.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical approval and Consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eStudy was approved by the institutional ethics committee of the Medical Faculty of the University of T\u0026uuml;bingen as well as the institutional data security and privacy review board and regional ethics committees\u003c/p\u003e\n\u003cp\u003eStudy was conducted in accordance with the first revision of the Declaration of Helsinki from 1975. Informed consent was obtained from all individual participants included in the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSeymour L, Bogaerts J, Perrone A, Ford R, Schwartz LH, Mandrekar S, et al. iRECIST: guidelines for response criteria for use in trials testing immunotherapeutics. The Lancet Oncology. 2017;18:e143-e52. doi:10.1016/S1470-2045(17)30074-8.\u003c/li\u003e\n\u003cli\u003ePersigehl T, Lennartz S, Schwartz LH. iRECIST: how to do it. Cancer imaging : the official publication of the International Cancer Imaging Society. 2020;20:2. doi:10.1186/s40644-019-0281-x.\u003c/li\u003e\n\u003cli\u003eBleker J, Kwee TC, Rouw D, Roest C, Borstlap J, de Jong IJ, et al. A deep learning masked segmentation alternative to manual segmentation in biparametric MRI prostate cancer radiomics. European radiology. 2022;32:6526-35. doi:10.1007/s00330-022-08712-8.\u003c/li\u003e\n\u003cli\u003ePaez D, Giammarile F, Orellana P. Nuclear medicine: a global perspective. Clinical and Translational Imaging. 2020;8:51-3. doi:10.1007/s40336-020-00359-z.\u003c/li\u003e\n\u003cli\u003eOECD. Health care utilization\u0026ndash;diagnostic exams,. https://statsoecdorg/indexaspx?queryid=30160. 2022.\u003c/li\u003e\n\u003cli\u003eSung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA: a cancer journal for clinicians. 2021;71:209-49. doi:10.3322/caac.21660.\u003c/li\u003e\n\u003cli\u003eSafiri S, Kolahi AA, Naghavi M, Global Burden of Disease Bladder Cancer C. Global, regional and national burden of bladder cancer and its attributable risk factors in 204 countries and territories, 1990-2019: a systematic analysis for the Global Burden of Disease study 2019. BMJ global health. 2021;6. doi:10.1136/bmjgh-2020-004128.\u003c/li\u003e\n\u003cli\u003eEbner R, Sheikh GT, Brendel M, Ricke J, Cyran CC. ESR Essentials: staging and restaging with FDG-PET/CT in oncology-practice recommendations by the European Society for Hybrid, Molecular and Translational Imaging. European radiology. 2024. doi:10.1007/s00330-024-11094-8.\u003c/li\u003e\n\u003cli\u003eEbner R, Sheikh GT, Brendel M, Ricke J, Cyran CC. ESR Essentials: role of PET/CT in neuroendocrine tumors-practice recommendations by the European Society for Hybrid, Molecular and Translational Imaging. European radiology. 2025;35:1903-12. doi:10.1007/s00330-024-11095-7.\u003c/li\u003e\n\u003cli\u003eAlotaibi NA, Yakar D, Glaudemans A, Kwee TC. Diagnostic errors in clinical FDG-PET/CT. Eur J Radiol. 2020;132:109296. doi:10.1016/j.ejrad.2020.109296.\u003c/li\u003e\n\u003cli\u003eToxopeus R, Kasalak \u0026Ouml;, Yakar D, Noordzij W, Dierckx R, Kwee TC. Is work overload associated with diagnostic errors on (18)F-FDG-PET/CT? European journal of nuclear medicine and molecular imaging.2024 Mar;51(4):1079-1084. doi:10.1007/s00259-023-6543-3.\u003c/li\u003e\n\u003cli\u003eYang F, Zamzmi G, Angara S, Rajaraman S, Aquilina A, Xue Z, et al. Assessing Inter-Annotator Agreement for Medical Image Segmentation. IEEE access : practical innovations, open solutions. 2023;11:21300-12. doi:10.1109/access.2023.3249759.\u003c/li\u003e\n\u003cli\u003eTran KA, Kondrashova O, Bradley A, Williams ED, Pearson JV, Waddell N. Deep learning in cancer diagnosis, prognosis and treatment selection. Genome medicine. 2021;13:152. doi:10.1186/s13073-021-00968-x.\u003c/li\u003e\n\u003cli\u003eLiao J, Li X, Gan Y, Han S, Rong P, Wang W, et al. Artificial intelligence assists precision medicine in cancer treatment. Frontiers in oncology. 2022;12:998222. doi:10.3389/fonc.2022.998222.\u003c/li\u003e\n\u003cli\u003eCiarmiello A, Giovannini E, Tutino F, Yosifov N, Milano A, Florimonte L, et al. Does FDG PET-Based Radiomics Have an Added Value for Prediction of Overall Survival in Non-Small Cell Lung Cancer? Journal of clinical medicine. 2024;13. doi:10.3390/jcm13092613.\u003c/li\u003e\n\u003cli\u003eShiyam Sundar LK, Yu J, Muzik O, Kulterer OC, Fueger B, Kifjak D, et al. Fully Automated, Semantic Segmentation of Whole-Body (18)F-FDG PET/CT Images Based on Data-Centric Artificial Intelligence. Journal of nuclear medicine : official publication, Society of Nuclear Medicine. 2022;63:1941-8. doi:jnumed.122.264063 [pii]264063 [pii]10.2967/jnumed.122.264063.\u003c/li\u003e\n\u003cli\u003eWasserthal J, Breit HC, Meyer MT, Pradella M, Hinck D, Sauter AW, et al. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiology Artificial intelligence. 2023;5:e230024. doi:10.1148/ryai.230024.\u003c/li\u003e\n\u003cli\u003eHasani N, Paravastu SS, Farhadi F, Yousefirizi F, Morris MA, Rahmim A, et al. Artificial Intelligence in Lymphoma PET Imaging:: A Scoping Review (Current Trends and Future Directions). PET Clin. 2022;17:145-74. doi: 10.1016/j.cpet.2021.09.006.\u003c/li\u003e\n\u003cli\u003eH\u0026auml;ggstr\u0026ouml;m I, Leithner D, Alv\u0026eacute;n J, Campanella G, Abusamra M, Zhang H, et al. Deep learning for [\u0026lt;sup\u0026gt;18\u0026lt;/sup\u0026gt;F]fluorodeoxyglucose-PET-CT classification in patients with lymphoma: a dual-centre retrospective analysis. The Lancet Digital Health. 2024;6:e114-e25. doi:10.1016/s2589-7500(23)00203-0.\u003c/li\u003e\n\u003cli\u003eGatidis S, Fr\u0026uuml;h M, Fabritius MP, Gu S, Nikolaou K, Foug\u0026egrave;re CL, et al. Results from the autoPET challenge on fully automated lesion segmentation in oncologic PET/CT imaging. Nature Machine Intelligence. 2024;6:1396-405. doi:10.1038/s42256-024-00912-9.\u003c/li\u003e\n\u003cli\u003eTarai S, Lundstr\u0026ouml;m E, Ahmad N, Strand R, Ahlstr\u0026ouml;m H, Kullberg J. Whole-body tumor segmentation from FDG-PET/CT: Leveraging a segmentation prior from tissue-wise projections. Heliyon. 2025;11:e41038. doi:https://doi.org/10.1016/j.heliyon.2024.e41038.\u003c/li\u003e\n\u003cli\u003eRahman WT, Wale DJ, Viglianti BL, Townsend DM, Manganaro MS, Gross MD, et al. The impact of infection and inflammation in oncologic (18)F-FDG PET/CT imaging. Biomedicine \u0026amp; pharmacotherapy = Biomedecine \u0026amp; pharmacotherapie. 2019;117:109168. doi:10.1016/j.biopha.2019.109168.\u003c/li\u003e\n\u003cli\u003eHess S, Scholtens AM, Gormsen LC. Patient Preparation and Patient-related Challenges with FDG-PET/CT in Infectious and Inflammatory Disease. PET Clin. 2020;15:125-34. doi:10.1016/j.cpet.2019.11.001.\u003c/li\u003e\n\u003cli\u003eMetser U, Even-Sapir E. Increased (18)F-fluorodeoxyglucose uptake in benign, nonphysiologic lesions found on whole-body positron emission tomography/computed tomography (PET/CT): accumulated data from four years of experience with PET/CT. Semin Nucl Med. 2007;37:206-22. doi:10.1053/j.semnuclmed.2007.01.001.\u003c/li\u003e\n\u003cli\u003eGatidis S, T K. A whole-body FDG-PET/CT dataset with manually annotated tumor lesions (FDG-PET-CT-Lesions) [Dataset]. The Cancer Imaging Archive. 2022. doi: 10.7937/gkr0-xv29 \u003c/li\u003e\n\u003cli\u003eGatidis S, Hepp T, Fruh M, La Fougere C, Nikolaou K, Pfannenberg C, et al. A whole-body FDG-PET/CT Dataset with manually annotated Tumor Lesions. Scientific data. 2022;9:601. doi:10.1038/s41597-022-01718-3 [pii]1718 [pii]10.1038/s41597-022-01718-3.\u003c/li\u003e\n\u003cli\u003eBoellaard R, Delgado-Bolton R, Oyen WJ, Giammarile F, Tatsch K, Eschner W, et al. FDG PET/CT: EANM procedure guidelines for tumour imaging: version 2.0. European journal of nuclear medicine and molecular imaging. 2015;42:328-54. doi:10.1007/s00259-014-2961-x.\u003c/li\u003e\n\u003cli\u003eRonneberger O. Medical image computing and computer-assisted intervention\u0026ndash;MICCAI 2015. Springer; 2015.\u003c/li\u003e\n\u003cli\u003eIsensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods. 2021;18:203-11. doi:10.1038/s41592-020-01008-z.\u003c/li\u003e\n\u003cli\u003eYu C, Anakwenze CP, Zhao Y, Martin RM, Ludmir EB, J SN, et al. Multi-organ segmentation of abdominal structures from non-contrast and contrast enhanced CT images. Scientific reports. 2022;12:19093. doi:10.1038/s41598-022-21206-3 [pii]21206 [pii]10.1038/s41598-022-21206-3.\u003c/li\u003e\n\u003cli\u003eXue H, Fang Q, Yao Y, Teng Y. 3D PET/CT tumor segmentation based on nnU-Net with GCN refinement. Phys Med Biol. 2023;68. doi:10.1088/1361-6560/acede6.\u003c/li\u003e\n\u003cli\u003eKalisch H, H\u0026ouml;rst F, K H, Kleesiek J, Seibold C. Autopet III challenge: Incorporating anatomical knowledge into nnUNet for lesion segmentation in PET/CT. arXiv preprint arXiv:221102701. 2024;2024.\u003c/li\u003e\n\u003cli\u003eAstaraki M, Bendazzolie S. Lesion Segmentation in Whole-Body Multi-Tracer PET-CT Images; a Contribution to AutoPET 2024 Challenge. arXiv preprint arXiv:221102701. 2024.\u003c/li\u003e\n\u003cli\u003eVirtanen P, Gommers R, Oliphant TE, Haberland M, Reddy T, Cournapeau D, et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nature methods. 2020;17:261-72. doi:10.1038/s41592-019-0686-2.\u003c/li\u003e\n\u003cli\u003eCardoso MJ, Li W, Brown R, Ma N, Kerfoot E, Wang Y, et al. Monai: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:221102701. 2022.\u003c/li\u003e\n\u003cli\u003eDice LR. Measures of the amount of ecologic association between species. Ecology. 1945;26:297-302.\u003c/li\u003e\n\u003cli\u003eSachpekidis C, Enqvist O, Ulen J, Kopp-Schneider A, Pan L, Jauch A, et al. Application of an artificial intelligence-based tool in [(18)F]FDG PET/CT for the assessment of bone marrow involvement in multiple myeloma. European journal of nuclear medicine and molecular imaging. 2023;50:3697-708. doi:10.1007/s00259-023-06339-5.\u003c/li\u003e\n\u003cli\u003eYousefirizi F, Klyuzhin IS, O JH, Harsini S, Tie X, Shiri I, et al. TMTV-Net: fully automated total metabolic tumor volume segmentation in lymphoma PET/CT images - a multi-center generalizability analysis. Eur J Nucl Med Mol Imaging. 2024;51:1937-54. doi: 10.007/s00259-024-6616-x. Epub 2024 Feb 8.\u003c/li\u003e\n\u003cli\u003eJemaa S, Fredrickson J, Carano RAD, Nielsen T, de Crespigny A, Bengtsson T. Tumor Segmentation and Feature Extraction from Whole-Body FDG-PET/CT Using Cascaded 2D and 3D Convolutional Neural Networks. Journal of digital imaging. 2020;33:888-94. doi:10.1007/s10278-020-00341-1.\u003c/li\u003e\n\u003cli\u003eHuang B, Chen Z, Wu PM, Ye Y, Feng ST, Wong CO, et al. Fully Automated Delineation of Gross Tumor Volume for Head and Neck Cancer on PET-CT Using Deep Learning: A Dual-Center Study. Contrast media \u0026amp; molecular imaging. 2018;2018:8923028. doi:10.1155/2018/8923028.\u003c/li\u003e\n\u003cli\u003eTeramoto A, Fujita H, Yamamuro O, Tamaki T. Automated detection of pulmonary nodules in PET/CT images: Ensemble false-positive reduction using a convolutional neural network technique. Medical physics. 2016;43:2821-7. doi:10.1118/1.4948498.\u003c/li\u003e\n\u003cli\u003eZincirkeser S, Sahin E, Halac M, Sager S. Standardized uptake values of normal organs on 18F-fluorodeoxyglucose positron emission tomography and computed tomography imaging. J Int Med Res. 2007;35:231-6. doi:10.1177/147323000703500207.\u003c/li\u003e\n\u003cli\u003eSharma P, Chatterjee P, Alvarado L, Dwivedi A. Standardized uptake value of normal organs on routine clinical [18F]FDG PET/CT: impact of tumor metabolism and patient-related factors. Nuclear Medicine Review. 2023;26:1-10. doi:10.5603/NMR.a2022.0036.\u003c/li\u003e\n\u003cli\u003eLindgren Belal S, Sadik M, Kaboteh R, Enqvist O, Ul\u0026eacute;n J, Poulsen MH, et al. Deep learning for segmentation of 49 selected bones in CT scans: First step in automated PET/CT-based 3D quantification of skeletal metastases. Eur J Radiol. 2019;113:89-95.\u003c/li\u003e\n\u003cli\u003eCarles M, Kuhn D, Fechter T, Baltas D, Mix M, Nestle U, et al. Development and evaluation of two open-source nnU-Net models for automatic segmentation of lung tumors on PET and CT images with and without respiratory motion compensation. Eur Radiol. 2024;34:6701-11. doi: 10.1007/s00330-024-10751-2. Epub 2024 Apr 25.\u003c/li\u003e\n\u003cli\u003eJafari E, Zarei A, Dadgar H, Keshavarz A, Manafi-Farid R, Rostami H, et al. A convolutional neural network-based system for fully automatic segmentation of whole-body [(68)Ga]Ga-PSMA PET images in prostate cancer. Eur J Nucl Med Mol Imaging. 2024;51:1476-87. doi: 10.007/s00259-023-6555-z. Epub 2023 Dec 14.\u003c/li\u003e\n\u003cli\u003eIm HJ, Bradshaw T, Solaiyappan M, Cho SY. Current Methods to Define Metabolic Tumor Volume in Positron Emission Tomography: Which One is Better? Nucl Med Mol Imaging. 2018;52:5-15. doi: 0.1007/s13139-017-0493-6. Epub 2017 Sep 19.\u003c/li\u003e\n\u003cli\u003eAndrearczyk V, Oreiller V, Boughdad S, Le Rest CC, Tankyevych O, Elhalawani H, et al. Automatic Head and Neck Tumor segmentation and outcome prediction relying on FDG-PET/CT images: Findings from the second edition of the HECKTOR challenge. Med Image Anal. 2023;90:102972. doi:S1361-8415(23)00232-3 [pii]10.1016/j.media.2023.102972.\u003c/li\u003e\n\u003cli\u003eOreiller V, Andrearczyk V, Jreige M, Boughdad S, Elhalawani H, Castelli J, et al. Head and neck tumor segmentation in PET/CT: The HECKTOR challenge. Med Image Anal. 2022;77:102336. doi:S1361-8415(21)00381-9 [pii]10.1016/j.media.2021.102336.\u003c/li\u003e\n\u003cli\u003eHatt M, Laurent B, Ouahabi A, Fayad H, Tan S, Li L, et al. The first MICCAI challenge on PET tumor segmentation. Med Image Anal. 2018;44:177-95. doi:S1361-8415(17)30189-5 [pii]10.1016/j.media.2017.12.007.\u003c/li\u003e\n\u003cli\u003eMoreau N, Rousseau C, Fourcade C, Santini G, Ferrer L, Lacombe M, et al. Influence of inputs for bone lesion segmentation in longitudinal (18)F-FDG PET/CT imaging studies. Annual International Conference of the IEEE Engineering in Medicine and Biology Society IEEE Engineering in Medicine and Biology Society Annual International Conference. 2022;2022:4736-9. doi:10.1109/EMBC48229.2022.9871081.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTables 1 to 3 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"ejnmmi-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejre","sideBox":"Learn more about [EJNMMI Research](http://ejnmmires.springeropen.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/ejre/default.aspx","title":"EJNMMI Research","twitterHandle":"@officialEANM","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Deep learning,nnU-Net,Segmentation,18F-FDG PET/CT","lastPublishedDoi":"10.21203/rs.3.rs-6895938/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6895938/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe number of [¹⁸F]fluorodeoxyglucose ([¹⁸F]FDG)-PET/CT scans performed has significantly increased in the last decade in line with the increasing trend of oncological malignancies. Such images, which signal high glucose-uptake areas are key in defining the extent of the disease, staging and response to therapy. Processing and evaluation of ([¹⁸F]FDG)-PET/CT scans, however, require manual annotation by well-trained specialists and above all time. In time and resource-constrained settings meeting the increasing demand for PET/CT scans has become challenging.\u003c/p\u003e\n\u003cp\u003eThe main goal of our study was to test the relationship between the volumes predicted by the deep learning algorithm and the manually segmented ones. The secondary objective goal was to measure the extent at which the predictive accuracy is associated with normal background uptake.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study sample included 1159 [¹⁸F]FDG-PET/CT scans from subjects with histologically confirmed diagnoses of lung cancer, lymphoma, and melanoma. 881 (70%) [¹⁸F]FDG-PET/CT scans were used as the training dataset and 232 (20%) scans were used as an internal validation dataset. A subsample of 116 (10%) [¹⁸F]FDG-PET/CT scans not used for training was used as the test dataset. The segmentation model was implemented with the nnU-Net convolutional network available in the MONAI framework. Model performance was measured with the Dice score. Correlation between manual and predicted segmentation was assessed using linear correlation. Totalsegmentator tool was used to identify lesions location and assess the tumor-to-background ratio (TBR) for quantitative analysis.\u003c/p\u003e\n\u003cp\u003eNetwork achieved Dice scores of 0.805 (validation) and 0.784 (test), showing strong agreement with manual segmentations. Anatomical localization was successful in 74% of the 7914 detected lesions. High correlation (R=0.88, p\u0026lt;0.0001) was observed between predicted and ground truth volumes. Segmentation accuracy improved with higher TBRs, as lesions with TBR\u0026gt;2 had significantly better Dice scores than those with lower contrast (TBR ≤ 1–2 or ≤1).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThese results are consistent with previous reports on PET-based segmentation, further validating nnU-Net as a reliable approach for detecting hypermetabolic lesions and assessing global disease burden in FDG-PET imaging. Moreover, the significant relationship between TBR and segmentation accuracy suggests the possibility of further improvements by integrating metabolic profile into the predictive model.\u003c/p\u003e","manuscriptTitle":"Global and regional accuracy of deep learning-based tumor segmentation from whole-body [¹⁸F]fluorodeoxyglucose PET/CT images","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-18 13:03:58","doi":"10.21203/rs.3.rs-6895938/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Major Revision","date":"2025-08-25T07:12:03+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"","date":"2025-07-18T17:47:49+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-07-13T20:15:21+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-06-26T09:21:27+00:00","index":"","fulltext":""},{"type":"submitted","content":"EJNMMI Research","date":"2025-06-21T08:29:00+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"ejnmmi-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejre","sideBox":"Learn more about [EJNMMI Research](http://ejnmmires.springeropen.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/ejre/default.aspx","title":"EJNMMI Research","twitterHandle":"@officialEANM","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"d1b69b93-16e6-4bb8-9771-0560925b781f","owner":[],"postedDate":"July 18th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-12-01T16:07:42+00:00","versionOfRecord":{"articleIdentity":"rs-6895938","link":"https://doi.org/10.1186/s13550-025-01333-4","journal":{"identity":"ejnmmi-research","isVorOnly":false,"title":"EJNMMI Research"},"publishedOn":"2025-11-28 15:57:20","publishedOnDateReadable":"November 28th, 2025"},"versionCreatedAt":"2025-07-18 13:03:58","video":"","vorDoi":"10.1186/s13550-025-01333-4","vorDoiUrl":"https://doi.org/10.1186/s13550-025-01333-4","workflowStages":[]},"version":"v1","identity":"rs-6895938","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6895938","identity":"rs-6895938","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}