A Deep Learning Lung Cancer Segmentation Pipeline to Facilitate CT-based Radiomics

A Deep Learning Lung Cancer Segmentation Pipeline to Facilitate CT-based Radiomics | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (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];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search A Deep Learning Lung Cancer Segmentation Pipeline to Facilitate CT-based Radiomics View ORCID Profile Alfred Chung Pui So , Daryl Cheng , Shahab Aslani , Mehran Azimbagirad , Daisuke Yamada , Roberta Dunn , Eleni Josephides , Ellie McDowall , Annie-Rose Henry , Andrea Bille , Nishanth Sivarasan , Adam Pennycuick , Eleni Karapanagiotou , Joseph Jacob doi: https://doi.org/10.1101/2025.06.17.25329213 Alfred Chung Pui So 1 Guy’s Cancer Centre, Guy’s and St Thomas’ NHS Foundation Trust , London, United Kingdom 3 Satsuma Lab, Hawke’s Institute, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Alfred Chung Pui So For correspondence: alfred.so{at}nhs.net Daryl Cheng 2 UCL Respiratory, University College London , London, United Kingdom 3 Satsuma Lab, Hawke’s Institute, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Shahab Aslani 3 Satsuma Lab, Hawke’s Institute, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Mehran Azimbagirad 3 Satsuma Lab, Hawke’s Institute, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Daisuke Yamada 3 Satsuma Lab, Hawke’s Institute, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Roberta Dunn 1 Guy’s Cancer Centre, Guy’s and St Thomas’ NHS Foundation Trust , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Eleni Josephides 1 Guy’s Cancer Centre, Guy’s and St Thomas’ NHS Foundation Trust , London, United Kingdom 4 School of Cancer and Pharmaceutical Sciences, King’s College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Ellie McDowall 1 Guy’s Cancer Centre, Guy’s and St Thomas’ NHS Foundation Trust , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Annie-Rose Henry 1 Guy’s Cancer Centre, Guy’s and St Thomas’ NHS Foundation Trust , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Andrea Bille 5 Department of Thoracic Surgery, Guy’s and St Thomas’ NHS Foundation Trust , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Nishanth Sivarasan 6 Department of Radiology, Guy’s and St Thomas’ NHS Foundation Trust , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Adam Pennycuick 2 UCL Respiratory, University College London , London, United Kingdom 7 Lungs for Living Research Centre, UCL Respiratory, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Eleni Karapanagiotou 1 Guy’s Cancer Centre, Guy’s and St Thomas’ NHS Foundation Trust , London, United Kingdom 4 School of Cancer and Pharmaceutical Sciences, King’s College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Joseph Jacob 3 Satsuma Lab, Hawke’s Institute, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Abstract Full Text Info/History Metrics Supplementary material Data/Code Preview PDF Abstract Background CT-based radio-biomarkers could provide non-invasive insights into tumour biology to risk-stratify patients. One of the limitations is laborious manual segmentation of regions-of-interest (ROI). We present a deep learning auto-segmentation pipeline for radiomic analysis. Patients and Methods 153 patients with resected stage 2A-3B non-small cell lung cancer (NSCLCs) had tumours segmented using nnU-Net with review by two clinicians. The nnU-Net was pretrained with anatomical priors in non-cancerous lungs and finetuned on NSCLCs. Three ROIs were segmented: intra-tumoural, peri-tumoural, and whole lung. 1967 features were extracted using PyRadiomics. Feature reproducibility was tested using segmentation perturbations. Features were selected using minimum-redundancy-maximum-relevance with Random Forest-recursive feature elimination nested in 500 bootstraps. Results Auto-segmentation time was ∼36 seconds/series. Mean volumetric and surface Dice-Sørensen coefficient (DSC) scores were 0.84 (±0.28), and 0.79 (±0.34) respectively. DSC were significantly correlated with tumour shape (sphericity, diameter) and location (worse with chest wall adherence), but not batch effects (e.g. contrast, reconstruction kernel). 6.5% cases had ‘missed’ segmentations; 6.5% required major changes. Pre-training on anatomical priors resulted in better segmentations compared to training on tumour-labels alone ( p <0.001) and tumour with anatomical labels ( p <0.001). Most radiomic features were not reproducible following perturbations and resampling. Adding radiomic features, however, did not significantly improve the clinical model in predicting 2-year disease-free survival: AUCs 0.67 (95%CI 0.59-0.75) vs 0.63 (95%CI 0.54-0.71) respectively ( p =0.28). Conclusion Our study demonstrates that integrating auto-segmentation into radio-biomarker discovery is feasible with high efficiency and accuracy. Whilst radiomic analysis show limited reproducibility, our auto-segmentation may allow more robust radio-biomarker analysis using deep learning features. 1. Introduction Lung cancer is the most common cause of cancer-related death worldwide [ 1 ]. Average life-expectancy in patients with lung cancer in the UK is less than 1-year, mainly due to diagnosis at an advanced stage [ 2 ]. Patients are increasingly diagnosed at an earlier stage, due to several factors including introduction of targeted lung cancer screening [ 2 ]. Furthermore, the treatment landscape of early-stage non-small cell lung cancers (NSCLC) is rapidly evolving with the additional use of chemoimmunotherapy and/or targeted therapy alongside surgery [ 3 ]. Despite this, recurrence rates can remain high. With more patients diagnosed earlier and living longer, it is important to balance the risk of overtreatment and quality of life. There are no routinely used clinical biomarkers for prognostication, with TNM staging remaining the gold standard. Promising results have been reported with novel molecular tests, ctDNA, digital pathology, and artificial-intelligence (AI) based models [ 4 – 7 ]. Currently, the implementation of these biomarkers remains limited by repeatability, resource-allocation, and time. Radiological biomarkers, or radio-biomarkers, are an attractive approach for clinicians because they can provide a non-invasive three-dimensional insight into cancer biology that cannot be visually evaluated by radiologists. Radiomics, a type of radio-biomarker, is a high-throughput computational method for quantitative medical imaging analysis using predefined mathematical features. NSCLC is one of the most studied oncological subtypes within radiomic research, but implementation of radiomic biomarkers has not reached clinical translation [ 8 , 9 ]. Previous studies have shown moderate correlation of radiomics with NSCLC prognosis, histology, actionable-genomic alterations (AGAs), and immunophenotypes [ 10 – 15 ]. However, radiomic features are not easily reproduced due to heterogeneity in clinical end-points, methodological transparency, image acquisition, pre-processing, segmentation quality, feature selection, and model classifier [ 16 – 18 ]. In addition, many radiomic studies do not incorporate robust assessments of feature stability [ 19 ]. Automatic segmentation, or auto-segmentation, has the potential to address some of these barriers including reducing manual segmentation time, avoiding observer variability, and standardising segmentation quality. To have a deployable radio-biomarker in the clinic, auto-segmentation is necessary to make these scalable. Several auto-segmentation models have already been previously published with variable performance [ 20 – 23 ]. However, integrating auto-segmentation within radio-biomarker studies has been less explored. Furthermore, the performance of these models will be influenced by the type of deep learning (DL) architecture used and training datasets. Using more stable regions-of-interest (ROI), such as whole lung, may be another attractive approach due to less heterogeneity in boundary definition with auto-segmentation algorithms readily available [ 24 ]. Most NSCLC-based radiomic research has been on patients receiving radiotherapy. Previous studies in resectable NSCLC have either focused only on stage 1 disease or collectively stage 1-3 [ 25 – 28 ]. It is important to distinguish these cohorts as resectable stage 2-3 NSCLC represents a distinct prognostic group. It is a heterogenous group with 5-year recurrence rates between 20-60%, representing a high-risk population where treatment escalation using chemotherapy with or without immunotherapy is recommended. The LACE meta-analysis demonstrated an absolute 4-5% benefit in 5-year overall survival and disease-free survival with the addition of platinum-based chemotherapy compared to surgery alone [ 29 ]. Neoadjuvant or peri-operative chemoimmunotherapy provides an additional 30-40% reduction in recurrence compared with chemotherapy, with two trials showing overall survival benefit [ 30 , 31 ]. Although significant breakthroughs have been made, some patients still experience early-recurrence. Current guidelines recommend 6-monthly surveillance CT scans following radical surgery for the first two years and yearly afterwards based on expert consensus [ 32 ]. Identifying early recurrence using radio-biomarkers can facilitate locoregional therapy or early introduction of systemic anti-cancer therapy (SACT) whilst the tumour-burden remains low. In this study, we present a research pipeline integrating semi-automated DL segmentation pretrained with anatomical priors to explore robust CT-based radiomic features in resectable stage 2-3 NSCLC. 2. Patients and methods This study complied with all relevant ethical regulations and received institutional approval. Informed consent was waived as this study uses a real-world database comprising routinely-collected, standard-of-care patient data and incorporates an opt-out consent process through Guy’s Cancer Cohort (Reference: 18/NW/0297) [ 33 ]. 2.1 Patients Patients were retrospectively identified from our tertiary cancer centre’s database from January 2015 to December 2024. All patients with resected stage 2A-3B NSCLC (TNM 9 th edition) were reviewed. 2-year disease-free survival (DFS) was used as the clinical endpoint. Exclusion criteria were: (1) absence of pre-operative baseline CT of the thorax, (2) follow-up time under 2-years, (3) death before recurrence, (4) inadequate information on histology and staging, and (5) pure or mixed neuroendocrine histology. 201 patients were assessed, of which 153 were eligible [Supplemental Fig. 1]. 2.2 Image acquisition and processing No cases were excluded based on CT acquisition protocols. 68.6% were intravenous contrast-enhanced scans and 57.5% used sharp convolution kernels. Average parameters include slice thickness 1.2mm, pixel size 0.76mm, and tube energy 113mA [Supplemental Table 1]. All images were resampled to isotropic voxel 1×1×1mm and set to lung window (level:-600HU; width: 1500HU). Images underwent grey-level normalisation 0-255 and fixed bin width discretisation to 25. 2.3 Image segmentation Primary tumours were auto-segmented using nnU-Net [ 34 ]. The nnU-Net was pretrained on 23 non-cancer containing cases with anatomical priors: airway, vessel, lung [ 35 ] [Supplemental Table 2]. The model was then finetuned on 50 cancer-containing cases derived from The Cancer Imaging Archive (TCIA) [ 36 – 40 ] [Supplemental Table 3]. All TCIA segmentations were reviewed and corrected as necessary. The hierarchical order of label input was lung, airway, vessel, tumour. This approach to model training mirrored how a radiologist would learn, first by understanding the thoracic structures in healthy and pathological lungs, then identifying lung cancers within their anatomical context [Supplemental Fig. 2]. This type of learning, like curriculum and anatomy-aware training, has been previously described but remains under researched in lung cancers [ 41 – 42 ]. The nnU-Net’s default hyperparameters and 3D configuration were used. All intra-tumoural segmentations were checked using a two-step review process and corrected on 3D-Slicer. The first review was performed by a medical oncology registrar with interest in thoracic oncology (A.C.P.S). The second review was performed by a thoracic radiologist with 10-years of experience (D.Y). Whole lungs were auto-segmented using TotalSegmentator [ 43 ]. Peri-tumoural segmentation was created by dilating the intra-tumoural segmentation by 3mm followed by removing the intra-tumoural region [ Fig. 1 ]. Although the 3mm dilation was chosen, there is currently no standardised definition of peri-tumoural region [ 44 ]. Download figure Open in new tab Figure 1: Segmentation pipeline. Both solid and non-solid components were considered as part of the segmentation. Peri-tumoural vasculature and distorted lung architecture were not included as part of the segmentation. During peri-tumoural dilation, the lung mask formed a boundary to prevent the dilation of the peri-tumoural segmentation into the thoracic wall and mediastinal structures. 2.4 Contour perturbations Perturbation methods, previously shown to be comparable to test-retest studies, were used to ensure feature repeatability [ 45 ]. Five perturbations (1mm) were performed: dilation, erosion, y-axis translation, x-axis translation, and contour randomisation. Gaussian noise perturbation was not included as the cohort already have heterogenous CT acquisition parameters. Perturbation methods were not performed on whole lung segmentations as there is less variation and would result in inclusion of non-lung segments. 2.5 Feature extraction and processing Radiomic features were extracted using PyRadiomics 3.1.0, an Image Biomarker Standardisation Initiative (IBSI) compliant software, and included the following filters: Laplacian of Gaussian, Wavelet, Square, SquareRoot, Logarithm, Exponential, Gradient, and LocalBinaryPattern3D [ 46 ]. A total 1967 radiomic features were extracted. Radiomic features were z-scaled. Multivariate Imputation by Chained Equations (MICE) was used to impute missing clinical data. AGAs, ethnicity, performance status, spread-through-air-space, and PDL1 data were excluded due to high missingness. One-hot and ordinal encoding was used. Continuous variables were z-scaled. 2.6 Radiomic feature selection No standardised feature selection approach exists, with each method offering different advantages and disadvantages [ 19 ]. In this study, a hybrid supervised feature selection approach of intraclass correlation coefficient (ICC) followed by minimum-redundancy-maximum-relevance recursive feature elimination-Random Forest (mRMR-RFE-RF) was used to account for feature repeatability, multicollinearity, and non-linear interactions [ Fig. 2 ]. We used the default RF hyperparameters on sklearn and set the RFE feature minimum to 10. This entire pipeline is nested within 500 bootstrap iterations to assess for feature stability. The performance of each bootstrap is aggregated for average model performance on the hold-out test set. Download figure Open in new tab Figure 2: Schematic representation of the study’s pipeline. After segmentation and review, intra-tumoural and peri-tumoural segmentations underwent five perturbations: dilation (1mm), erosion (1mm), y-axis translation (1mm), x-axis translation (1mm), and contour randomisation (1mm). Features were extracted within each perturbation using PyRadiomics with the application of filters for higher-order features. Features were initially filtered based on robustness against perturbations using a conservative ICC of ≥0.75. An ICC of 0.75-0.9 and ≥0.9 indicates good and excellent repeatability respectively. Following initial filtering, features with near-zero variance were removed. Features that remained then underwent mRMR, a pairwise filter method that maximises relevance and removes redundancy through mutual information. It is used as our first step to reduce the feature pool of highly redundant features before further feature selection. A pseudo-elbow method was used which plotted the average mRMR feature importance scores against number of features added, allowing a more data-driven selection. The next step involved using RFE-RF, a feature selection wrapper method that trains a RF model and progressively removes features until the final desired number of features remain. Predictive importance of radiomic features is tested through 500 bootstrap iterations. The most common radiomic features, present across ≥50% of bootstrap iterations, are combined with clinical features and undergo further LASSO feature reduction before model building. ICC, intraclass correlation coefficient. LASSO, least absolute shrinkage and selection operator. mRMR, minimum-redundancy-maximum-relevance. RFE-RF, recursive feature elimination-Random Forest. Figure created with BioRender. 2.7 Clinical feature selection The estimated maximum events per predictor (EPP) was 7 using Riley et al’s 2020 method (sample size 153, outcome proportion 0.49, R 2 CS 0.15, 8 0.05) [ 47 ]. This restriction was not relevant to RF as it is less reliant on the traditional EPP-based restrictions. Clinical features were selected through a combination of clinical features of interest, assessment of multicollinearity, and backward feature elimination. The selected features were trained using logistic regression across 500 bootstraps. 2.8 Combined clinical-radiomic model Combining clinical with radiomic features can be done at an early or late stage of a feature selection pipeline. Although early fusion is advantageous as it considers interactions between clinical and radiomic features early on, RF is often biased towards selecting continuous over categorical features based on how it calculates feature importance due to splitting opportunities [ 48 ]. To address this, clinical and radiomic features were combined after their individual selection process. To limit the number of features utilised (EPP=7), we used least absolute shrinkage and selection operator (LASSO) to reduce the feature pool. 2.8 Statistical analysis The performance of the auto-segmentation model was evaluated using volumetric DSC (vDSC), Intersection over Union (IoU), sensitivity, specificity, precision, and accuracy. Hausdorff distance was not calculated as it can be confounded by presence of synchronous lesions. Surface DSC (sDSC) was reported as it more closely predicts time-saved in manual segmentation compared to overlap metrics [ 49 ]. A 1mm sDSC tolerance threshold was used to reflect a strict threshold within the maximum boundary variation introduced by perturbations. The final segmentations used within the prediction model was considered the ground-truth as it had been reviewed by the thoracic radiologist who has been blinded to this study. Spearman rank and Mann-Whitney U were used to correlate independent variables with DSC scores. Kruskal-Wallis followed by Dunn’s post-hoc test was used for multiple comparisons between tumour location and DSC scores. Beta-regression was used for multivariate analysis in DSC score evaluation. Wilcoxon-rank was used to compare different auto-segmentation performance. Chi-squared and Mann-Whitney U were used to compare patient characteristics. The performance of the predictive model was evaluated by Receiver-Operator Characteristic (ROC) curves: Area Under the Curve (AUC), F1 score, specificity, sensitivity, recall, precision, accuracy, and brier calibration score. Probability threshold for classification was kept at default 0.5. ROC curves were compared using DeLong’s test. As this is an exploratory study, the alpha significance threshold was kept at 0.05. All analysis was performed on Python 3.10.12: matplotlib 3.10.0, MLstatkit 0.1.7, mrmr_selection 0.2.8, nibabel 5.3.2, numpy 2.1.3, pingouin 0.5.5, pyroc 0.1.1, scikit-learn 1.6.1, scikit-optimize 0.10.2, scipy 1.15.1, SimpleITK 2.4.1, statsmodels 0.14.4, and surface-distance 0.1. 3. Results 3.1 Patient characteristics 153 patients were included in this study with a median age of 69 years old (range 46-88), 43.8% male sex, 75.8% non-squamous cell histology, and 56.2% stage 3. 88.2% of patients had a lobectomy. 7.2% and 50.3% of patients received neoadjuvant and adjuvant SACT respectively [ Table 1 ]. The 2-year DFS rate following surgical treatment was 49.0% ( n =75/153). Of those that recurred within 2 years, the median time-to-recurrence was 252 days (95%CI 221-325 days) [Supplemental Fig. 3]. View this table: View inline View popup Table 1: Patient baseline demographics, cancer-specific characteristics, and treatment regimens received by patients. Percentages are displayed as within column percentages. The p-value was calculated using Chi-squared test for categorical variables and Mann-Whitney U test for continuous variables. 3.2 Auto-segmentation performance The average auto-segmentation time per patient case was 36 seconds (range 12-91), equating to ∼7.7 slices per second. This includes segmenting the tumour, whole lungs, airways, and vasculature. Our model had good performance across all metrics with mean vDSC, sDSC, IoU, sensitivity, specificity, precision, and recall of 0.84 (±0.28), 0.79 (±0.34), 0.80 (±0.31), 0.89 (±0.28), 1.00 (±<0.01), 0.86 (±0.26), and 1.00 (±<0.01) respectively [ Fig. 3 ]. In total, there were 6.5% (n=10/153) ‘missed’ segmentations and 6.5% (n=10/153) segmentations that required major changes. Majority of these tumours were adherent to extra-pulmonary structures or had parenchymal distortion [Supplemental Fig. 4, Supplemental Table 4]. Download figure Open in new tab Figure 3: Box-whiskers plot reporting standard metrics for the pre-trained auto-segmentation model (model_A). ACC, accuracy. DSC, Dice-Sørensen coefficient. DSC_surf, surface DSC. DSC_vol, volumetric DSC. IoU, Intersetion over Union. PREC, precision. SEN, sensitivity. SPEC, specificity. 3.3 Tumour and Batch factors on auto-segmentation performance Six tumour factors (volume, sphericity, maximum diameter, surface:volume ratio, semi-solid component, location) and four batch factors (contrast, original slice thickness, original pixel size, kernel type) were assessed against segmentation performance. Median tumour volume and maximum 3D diameter were 10727mm 3 (IQR 3152-21140) and 49.4mm (IQR 30.0-75.0) respectively. 20.9% and 13.7% of tumours were adhering to the chest wall and mediastinum respectively [Supplemental Table 5]. 82.4% were pure solid tumours. On univariate analysis, both DSC scores were positively correlated with sphericity, pure solid tumours, smooth kernel, and presence of intravenous contrast [Supplemental Table 6]. Larger tumour volume and diameter had a weak inverse relationship with performance. On Dunn’s multiple comparisons, vDSC and sDSC scores were worse with tumours adhering to the hilum ( p =0.003; p =0.001), apices ( p =0.028; p =0.006), and chest wall ( p =0.048; p =0.16) [Supplemental Fig. 5]. On multivariate beta regression, sphericity, maximum 3D diameter, and peripheral invasion (i.e. tumour adherence to chest wall, apices, diaphragm) remained independently correlated with performance [ Table 2 ]. Involvement of the hilum or mediastinum were trending towards worse segmentation but did not reach statistical significance. Although segmentation of semi-solid nodules was comparable to pure solid nodules, visual inspection on the poorly segmented semi-solid nodules (DSC <0.6) showed that the non-solid component was commonly missed [Supplemental Fig. 4]. None of the batch factors were independently correlated with segmentation performance. All variables had a variance inflation factor (VIF) <2. View this table: View inline View popup Download powerpoint Table 2: Multivariate beta regression analysis of factors impacting volumetric and surface DSC scores. Multiple linear regression (e.g. ordinary least squares) was not appropriate due to bounded outcome (DSC score 0-1), asymmetric distribution, and heteroskedasticity. Sphericity measures the roundness of the tumour region relative to a sphere, and surface:volume ratio measures the general compactness of the shape relative to the volume (a lower score indicates more sphere-like shape). Pixel size and slice thickness represent original sizes before pre-processing (all images were resampled to isotropic voxel 1×1×1mm). Isolated tumours refer to tumours not in contact with other major boundary structures (e.g. chest wall, hilum, mediastinum). Volumetric DSC: Pseudo R 2 0.320. Surface DSC: Pseudo R 2 0.348. S:V ratio, surface:volume ratio. *Central invasion refers to lesion adhering to the mediastinum (n=21/153) or hilum (n=10/153). ** Peripheral invasion refers to lesion adhering to the chest wall (n=32/153), apices (n=7/153), or diaphragm (n=5/153). 3.4 Dataset influence on auto-segmentation performance To test the influence of pre-training and anatomical labels on model performance, we compared the pre-trained model (model_A) against nnU-Net models trained only on tumour-labels (model_B) and tumour-airway-vessel-lung labels (model_C) with the same TCIA dataset. Model_A performed significantly better than model_B and _C across vDSC, sDSC, IoU score, and sensitivity ( p <0.001) [ Table 3 ; Supplemental Fig. 6]. The addition of anatomical labels with tumour labels during training (model_C) improved all metrics scores compared to training with tumour labels alone (model_B), but only accuracy and precision reached statistical significance. Compared to both models, model_A had less missed cases and poorly segmented cases (vDSC <0.6). Interestingly, not all poorly segmented cases overlapped between models [Supplemental Fig. 7]. View this table: View inline View popup Table 3: Wilcoxon-rank pairwise comparison of auto-segmentation metrics across models. Model_A: pre-trained using anatomical priors (lung, airway, vessel) on cases without lung tumours, then finetuned on cases with lung cancers. Model_B: no pre-training, only trained on tumour labels. Model_C: no pre-training, trained on tumour labels as well as anatomical labels (lung, airway, vessel). IoU, intersection over union. sDSC, surface DSC. std, standard deviation. vDSC, volumetric DSC. On reverse validation, we employed the same transfer learning approach with anatomical priors using the GCC dataset (n=153) and validating on the unseen TCIA dataset (n=50). The pre-trained model (model_D) performed marginally better than the model trained on tumour labels alone (model_E) across all metrics but did not reach statistical significance [Supplemental Table 7; Supplemental Fig. 8]. 3.5 Correlation of clinical features with 2-year DFS For the clinical model, 7 clinical features were selected: staging, R status, sex, PET-SUV-max of primary tumour, reception of SACT, pleural invasion, and type of surgical treatment [Supplemental Table 8]. On multivariate analysis, only staging and sex were significantly correlated with 2-year DFS. Interaction analysis demonstrated that reception of SACT in stage 3 disease reduced probability of recurrence from 0.72 to 0.52 (marginal effect dy/dx = –0.32; p =0.076). LASSO feature reduction similarly identified staging, sex, and PET-SUV-max as prognostic features [Supplemental Fig. 9]. The multivariate model, either using features from the classic approach or LASSO, had similar predictive performance with AUCs 0.63 (95%CI 0.54-0.71) and 0.61 (95%CI 0.52-0.70) respectively [Supplemental Table 9; Supplemental Fig. 10]. Using a RF classifier did not improve performance, suggesting absence of significant non-linear interactions. 3.6 Intra-tumour, peri-tumoural, and whole lung radiomic models On average, 55-60% and 20-25% of radiomic features achieved an ICC ≥0.75 and ≥0.90 across multiple perturbations. Although 150-200 features contribute >50% of total feature relevance score during mRMR, selecting the top 100 features achieved the most consistent performance. Median number of features selected across bootstraps following mRMR-RFE-RF for intra-tumoural, peri-tumoural, and whole lung radiomics were 61 (IQR 32-83), 56 (IQR 30-80), and 50 (IQR 34-78) respectively. Intra-tumoural, peri-tumoural, and whole lung radiomics had wide confidence intervals, suggesting radiomic feature instability, with AUCs 0.63 (95%CI 0.38-0.84), 0.56 (0.31-0.77), and 0.58 (95%CI 0.34-0.78) respectively [Supplemental Fig. 11]. Limiting feature count between 10-30 did not improve AUC scores. Nine intra-tumoural, 7 peri-tumoural, and 2 whole lung features were present in ≥50% of bootstrap iterations [Supplemental Table 10]. 3.7 Clinical-radiomics model Adding radiomic features from a single ROI source did not improve predictive performance. Combining the common intra-tumoural, peri-tumoural, and whole lung features with the clinical model followed by LASSO feature reduction improved overall performance with AUC of 0.67 (95%CI 0.59-0.75) and Brier score of 0.23 (95%CI 0.21-0.25) [ Table 4 , Supplemental Fig. 12]. However, this was not statistically significant compared to the clinical model (ΔAUC 0.04; z-statistic 1.08, p =0.281). The use of an RF classifier did not improve model performance. View this table: View inline View popup Download powerpoint Table 4: Clinical-radiomic model performance and selected feature combinations. Coefficients for the logistic regression model were: staging (3>2; +0.46), primary surgical treatment (sub-lobar resection>lobectomy;-0.49), sex (male>female; +0.51), tumour_log-sigma-3-0-mm-3D_firstorder_Skewness (-0.62), tumour_log-sigma-4-0-mm-3D_glcm_ClusterShade (-0.81), peri_log-sigma-4-0-mm-3D_firstorder_Skewness (+0.53), lung_wavelet-HHL_glcm_MCC (-0.33). 3.8 Interpreting radiomic features Intra-tumoural ‘skewness’ and ‘cluster shade’ were inversely correlated with recurrence, whereas peri-tumoural ‘skewness’ was positively correlated with recurrence. ‘Skewness’ is a first order feature that represents asymmetry of the distribution of intensity along the global mean, whilst ‘cluster shade’ is a textural feature that represents asymmetry or non-uniformity of co-occurring pixels. Pearson correlation demonstrated that none of the commonly selected radiomic features have a strong correlation with tumour size, PET-SUV-max, presence of lvi, pleural invasion, and histology. Exploratory analysis suggests a moderate positive correlation of predominant micropapillary growth pattern with intra-tumoural ‘cluster shade’ ( r =0.46-0.47; p <0.001) [Supplemental Fig. 12]. 4.0 Discussion This study demonstrated that a semi-automated DL segmentation pipeline for radio-biomarker discovery is feasible with high efficiency and good segmentation performance. The addition of pre-training using anatomical priors improved auto-segmentation performance compared to training with tumour labels alone and training with anatomical labels (without pretraining). The accuracy of the nnU-Net model was not influenced by batch effects (e.g contrast, reconstruction kernel) but rather by tumour shape and location. In general, more irregularly shaped tumours and tumours adhering to the chest wall had lower DSC scores. This study also showed that within this cohort of resected stage 2A-3B NSCLCs, CT-based traditional handcrafted radiomic features were highly unstable with poor repeatability. In recent years, there has been growing interest in using auto-segmentation in radio-biomarker studies. This is in part due to reduction in manual segmentation time as well as enhancing consistency and reproducibility. Furthermore, auto-segmentation provides the opportunity for radio-biomarkers to be deployable in clinical practice due to scalability. Several auto-segmentation models have been published with moderate to good performance [ 20 – 23 ]. For example, Primakov et al (2022) used a modified 2D CNN trained on nearly 1000 CT scans and achieved a vDSC and sDSC of 0.82±0.17 and 0.63±0.28 respectively [ 20 ]. In that study, higher tumour complexity, as measured by the necessity of PET-guided manual segmentation or tumour adherence to mediastinum or chest wall, were associated with poorer segmentations [ 20 ]. The nnU-Net is an attractive approach to segmentation as it is self-configuring and easy to adapt. Currently, evidence in integrating auto-segmentation within radio-biomarker development in lung cancer remains limited. Furthermore, surface agreement (e.g. sDSC), rather than overlap (e.g. vDSC, IoU), is more representative of time-saved from contouring but is usually not reported [ 49 ]. Our auto-segmentation model has comparable, if not better, performance compared to most studies with vDSC, sDSC, and IoU of 0.84 (±0.28), 0.79 (±0.33), and 0.80 (±0.31). Furthermore, this was achieved within stage 2-3 NSCLCs of various radiological phenotypes and CT acquisition parameters. The reason segmentation performance was not affected by batch effects was likely due to the heterogenous training dataset and nnU-Net’s self-calibration during training. The segmentations that were missed or required major revisions correlated with more irregular tumour shapes and adherence to the chest wall. Unfortunately, it is difficult to train complex tumour shapes as there is no clear agreement amongst expert clinicians regarding ground-truth boundaries [ 50 ]. On the other hand, there may be several approaches to improve auto-segmentation of tumours adhering to major boundary structures. These include increasing anatomical priors (e.g. chest wall, heart), excluding pixels in non-lung regions, limiting the input volume by cropping around the tumour region, and adding boundary streams to highlight edge information [ 51 – 53 ]. Medical image segmentation is usually considered a network architecture and/or data problem. One of the major barriers is the availability of large datasets with high quality segmentations. In our study, the combination of pre-training with three anatomical priors (lung, airway, vessel) on non-cancerous lungs followed by finetuning on tumour labels improved segmentation performance despite a small training dataset. It is most likely that pre-training provided additional anatomical context and created bounding regions to localise the tumour. Using anatomical priors have been previously shown to improve segmentations on heart, brain, liver, pancreas, and lymph nodes [ 52 , 54 – 57 ]. This approach has been less explored in lung tumours. Mathai et al (2025) trained two nnU-Net models, one without anatomical priors and one with 28 anatomical priors, for lung nodule segmentation [ 58 ]. Although incorporating 28 anatomical priors reduced vDSC score from 0.75 (±0.19) to 0.73 (±0.18), it managed to localise all false positives within the lung parenchyma unlike the baseline model [ 58 ]. This suggests a mechanism of constraining predictions within anatomical regions through priors [ 58 ]. However, using 28 anatomical priors may be adding too much noise. Asha et al (2024) utilised a transfer learning approach using the Segment Anything Model to leverage bounding box prompts for segmenting the LIDC-IDRI cohort (predominantly lung nodules) [ 59 ]. Although not exactly using anatomical priors, this approach achieved the best vDSC score of 0.97 compared to other DL networks [ 59 ]. It is important to note that these studies focussed primarily on lung nodules. There remains significant interest in exploring radio-biomarkers for lung cancer as it is non-invasive and allows 3D spatial analysis of the tumour. Furthermore, compared to other-omics (e.g. genomics, transcriptomics, computational pathology) radio-biomarkers are cheaper and faster to develop for clinical use. Previous radiomic research in resectable NSCLC has primarily focused on stage 1 disease or combining stage 1-3 [ 25 – 28 ]. In our study, the combined clinical-radiomic model performed marginally better in predicting 2-year DFS compared to unimodal datasets but was not statistically significant. Staging remained the strongest predictor of recurrence in this cohort with a treatment interaction effect observed. Recurrence was weakly correlated with intra-tumoural ‘skewness’ and ‘cluster shade’, and peri-tumoural ‘cluster shade’. Given the weak predictive signal, these results need to be interpreted with caution. We found that most radiomic features were not robust to perturbation experiments and only 20-25% have an excellent concordance (ICC ≥0.9). Bootstrapping was wrapped around the feature selection method, mRMR-RFE-RF, to provide a realistic view of model stability and demonstrated that majority of features are not repeatable. Most radiomic studies do not integrate ICC and bootstrapping within their pipeline, which may explain why many radiomic signatures have not been reproduced. Batch effects are known to impact traditional radiomic features and this barrier will continue to exist [ 60 , 61 ]. Whilst older harmonisation techniques may be used, there is increasing promise in more flexible DL approaches [ 62 ]. Alternatively, developing batch-invariant radio-biomarkers is another option (e.g. shape). Traditional shape features are usually not prognostic and novel approaches need to be considered. Wu et al (2021) previously defined four prognostic radiological subtypes of NSCLC clustered based on tumour volume, regional margin variation, shape symmetry, and shape regularity [ 63 ]. Peri-tumoural vascular morphology is another promising but under investigated radio-biomarker [ 64 ]. The emerging role of anti-angiogenic drugs in both AGA and non-AGA advanced NSCLCs highlight the importance of this mechanism. The use of deep features has also been increasing. For example, deep features derived from 3D CNN in stage 1-3 NSCLCs was correlated with genetic markers of cellular proliferation [ 65 ]. Radio-biomarkers, however, cannot fully encapsulate the complex tumour microenvironment compared to histopathology and genomics. In the era of AI, integration of multi-modal datasets will be the future. However, the role of radio-biomarkers remains to be defined. It is likely that traditional radiomics is limited by the voxel resolution of clinical-grade CTs. The use of high-resolution imaging, such as micro-CT, has yet to be explored. Although micro-CT is currently limited to human ex-vivo scanning, it may have a role in post-operative risk-stratification in early-stage NSCLCs. Our study has several limitations. First, although our sample size compared favourably with the literature for this type of lung cancer cohort, it remains small and potentially underpowered. Second, we excluded death from the primary endpoint. Due to high missingness in cause of death we could not accurately identify cancer-specific mortality. By excluding these patients, this reduced our sample size. Third, batch effects could have impacted radiomic stability. Although ComBat harmonisation is available, it requires the assumption that batches have similar biological variance and that differences are driven by technical factors [ 66 ]. In this cohort, due to the range of tumour shapes and invasiveness, ComBat harmonisation could mask biological signals. To adjust for this, we explicitly evaluated radiomic repeatability through these variations using perturbations and bootstrapping. Finally, there was significant treatment heterogeneity with regards to whether a patient received SACT and the type of SACT given. Although SACT was not significantly associated with recurrence in our study, several phase 3 randomised control trials have proven its benefit. These differences could represent the underpowering of our study. However, a stage-dependent effect was recognised during interaction analysis. 5.0 Conclusion In summary, we developed a semi-automated pipeline incorporating deep learning auto-segmentation to explore robust radiomic biomarkers in resected stage 2A-3B NSCLC. Our findings show that auto-segmentation can be effectively integrated within radio-biomarker discovery with an acceptable error margin. Although it is possible to identify robust radiomic biomarkers using this approach, the overall predictive signal remains weak and the majority of radiomic features lack reproducibility. The development of batch-invariant radio-biomarkers is vital for the consideration of future clinical adoption within oncology. Integrating auto-segmentation pipelines during biomarker developments will make these approaches clinically scalable. Data Availability Due to the potentially identifiable nature of the raw data regarding our patients, this information will remain confidential and not shared. The TCIA datasets and nnU-Net are open-source, and can be found on the following websites: https://www.cancerimagingarchive.net/ & https://github.com/MIC-DKFZ/nnUNet . Footnotes CRediT authorship contribution statement Alfred Chung Pui So: Conceptualisation, Methodology, Software, Validation, Formal analysis, Investigation, Data Curation, Writing – Original Draft, Visualisation, Funding Acquisition. Daryl Cheng: Methodology, Formal analysis, Writing – Review & Editing. Shahab Aslani: Methodology, Formal analysis, Writing – Review & Editing. Mehran Azimbagirad: Methodology, Software, Writing – Review & Editing. Daisuke Yamada: Methodology, Validation, Investigation, Writing – Review & Editing. Roberta Dunn: Data Curation. Ellie McDowall: Data Curation. Annie-Rose Henry: Data Curation. Andrea Billie: Data Curation, Resources. Nishanth Sivarasan: Conceptualisation, Data Curation. Adam Pennycuick: Conceptualisation, Methodology, Writing – Review & Editing, Supervision. Eleni Karapanagiotou: Conceptualisation, Methodology, Resources, Writing – Review & Editing, Supervision, Funding Acquisition. Joseph Jacob: Conceptualisation, Methodology, Resources, Writing – Review & Editing, Supervision, Funding Acquisition. Data statement Due to the potentially identifiable nature of the raw data regarding our patients, this information will remain confidential and not shared. The TCIA datasets and nnU-Net are open-source, and can be found on the following websites: https://www.cancerimagingarchive.net/ & https://github.com/MIC-DKFZ/nnUNet . Funding This work was supported by the Pan-London Cancer Alliance Fellowship; the Wellcome Trust (209553/Z/17/Z and 227835/Z/23/Z); and Chan Zuckerberg Initiative Foundation (CZIF2024-009938). The funders did not play any role in the conception, design, conduct, or reporting of this study. Declaration of competing interest ACPS declares honorarium from Roche. JJ declares consultancy fees from Boehringer Ingelheim, F. Hoffmann-La Roche, GlaxoSmithKline, NHSX; fees from advisory boards for Boehringer Ingelheim, F. Hoffmann-La Roche; lecture fees from Boehringer Ingelheim, F. Hoffmann-La Roche, Takeda; grant funding from GlaxoSmithKline, Wellcome Trust, Microsoft Research, Gilead Sciences, Chan Zuckerberg Initiative and UK patent application numbers: GB2113765.8 and GB2211487.0. All other authors involved declare no relevant conflicts of interest. Acknowledgements We would like to thank all the patients whose data was used for this study. We would also like to acknowledge the Pan-London Cancer Alliance Fellowship, the Wellcome Trust, and Chan Zuckerberg Initiative Foundation whose funding allowed this research to be completed. JJ was also supported by the NIHR UCLH Biomedical Research Centre. Ethics statement This study received ethical and institutional approval from the Guy’s Cancer Cohort research ethics committee. Patient informed consent was waived as this study uses a real-world database comprising of routinely-collected, standard-of-care patient data and incorporates an opt-out consent process through Guy’s Cancer Cohort research database (Reference: 18/NW/0297). 6. References 1. ↵ Zhou J , Xu Y , Liu J , Feng L , Yu J , Chen D . Global burden of lung cancer in 2022 and projections to 2050: Incidence and mortality estimates from GLOBOCAN . Cancer Epidemiol . 2024 Dec ; 93 : 102693 . doi: 10.1016/j.canep.2024.102693 . PMID: 39536404 . OpenUrl CrossRef PubMed 2. ↵ Royal College of Surgeons of England. National Lung Cancer Audit State of the Nation 2024. London: NLCA; 2024 . Available from: https://www.lungcanceraudit.org.uk/wp-content/uploads/2024/05/NLCA-State-of-the-Nation-2024_16.05.24_V2.0.pdf . Accessed on 2025 May . 3. ↵ Spicer JD , Cascone T , Wynes MW , Ahn MJ , Dacic S , Felip E , Forde PM , Higgins KA , Kris MG , Mitsudomi T , Provencio M , Senan S , Solomon BJ , Tsao MS , Tsuboi M , Wakelee HA , Wu YL , Chih-Hsin Yang J , Zhou C , Harpole DH , Kelly KL . Neoadjuvant and Adjuvant Treatments for Early Stage Resectable NSCLC: Consensus Recommendations From the International Association for the Study of Lung Cancer . J Thorac Oncol . 2024 Oct ; 19 ( 10 ): 1373 – 1414 . doi: 10.1016/j.jtho.2024.06.010 . PMID: 38901648 . OpenUrl CrossRef PubMed 4. ↵ Black JRM , Bartha G , Abbott CW , Boyle SM , Karasaki T , Li B , Chen R , Harris J , Veeriah S , Colopi M , Bakir MA , Liu WK , Lyle J , Navarro FCP , Northcott J , Pyke RM , Hill MS , Thol K , Huebner A , Bailey C , Colliver EC , Martínez-Ruiz C , Grigoriadis K , Pawlik P , Moore DA , Marinelli D , Shutkever OG , Murphy C , Sivakumar M; TRACERx consortium; Shaw JA, Hackshaw A, McGranahan N, Jamal-Hanjani M, Frankell AM, Chen RO , Swanton C. Ultrasensitive ctDNA detection for preoperative disease stratification in early-stage lung adenocarcinoma. Nat Med . 2025 Jan ; 31 ( 1 ): 70 – 76 . doi: 10.1038/s41591-024-03216-y . PMID: 39806071 . OpenUrl CrossRef PubMed 5. Biswas D , Liu YH , Herrero J , Wu Y , Moore DA , Karasaki T , Grigoriadis K , Lu WT , Veeriah S , Naceur-Lombardelli C , Magno N , Ward S , Frankell AM , Hill MS , Colliver E , de Carné Trécesson S , East P , Malhi A , Snell DM , O’Neill O , Leonce D , Mattsson J , Lindberg A , Micke P , Moldvay J , Megyesfalvi Z , Dome B , Fillinger J , Nicod J , Downward J , Szallasi Z; TRACERx Consortium; Hackshaw A, Jamal-Hanjani M, Kanu N, Birkbak NJ , Swanton C. Prospective validation of ORACLE, a clonal expression biomarker associated with survival of patients with lung adenocarcinoma. Nat Cancer . 2025 Jan ; 6 ( 1 ): 86 – 101 . doi: 10.1038/s43018-024-00883-1 . PMID: 39789179 . OpenUrl CrossRef PubMed 6. Kludt C , Wang Y , Ahmad W , Bychkov A , Fukuoka J , Gaisa N , Kühnel M , Jonigk D , Pryalukhin A , Mairinger F , Klein F , Schultheis AM , Seper A , Hulla W , Brägelmann J , Michels S , Klein S , Quaas A , Büttner R , Tolkach Y . Next-generation lung cancer pathology: Development and validation of diagnostic and prognostic algorithms . Cell Rep Med . 2024 Sep 17; 5 ( 9 ): 101697 . doi: 10.1016/j.xcrm.2024.101697 . PMID: 39178857 . OpenUrl CrossRef PubMed 7. ↵ Kim PJ , Hwang HS , Choi G , Sung HJ , Ahn B , Uh JS , Yoon S , Kim D , Chun SM , Jang SJ , Go H . A new model using deep learning to predict recurrence after surgical resection of lung adenocarcinoma . Sci Rep . 2024 Mar 16; 14 ( 1 ): 6366 . doi: 10.1038/s41598-024-56867-9 . PMID: 38493247 . OpenUrl CrossRef PubMed 8. ↵ Pan F , Feng L , Liu B , Hu Y , Wang Q . Application of radiomics in diagnosis and treatment of lung cancer . Front Pharmacol . 2023 Nov 1; 14 : 1295511 . doi: 10.3389/fphar.2023.1295511 . PMID: 38027000 . OpenUrl CrossRef PubMed 9. ↵ Fornacon-Wood I , Faivre-Finn C , O’Connor JPB , Price GJ . Radiomics as a personalized medicine tool in lung cancer: Separating the hope from the hype . Lung Cancer . 2020 Aug ; 146 : 197 – 208 . doi: 10.1016/j.lungcan.2020.05.028 . PMID: 32563015 . OpenUrl CrossRef PubMed 10. ↵ Lin H , Hua J , Wang Y , Chen M , Liang Y , Yan L , Zhao W , Luo S , Hong D , Chen X , Pan X , Liu J , Liu Z . Prognostic and predictive values of a multimodal nomogram incorporating tumor and peritumor morphology with immune status in resectable lung adenocarcinoma . J Immunother Cancer . 2025 Mar 6; 13 ( 3 ): e010723 . doi: 10.1136/jitc-2024-010723 . PMID: 40050046 . OpenUrl CrossRef PubMed 11. Jimenez Londoño GA , Pérez-Beteta J , Amo-Salas M , Honguero-Martinez AF , Pérez-García VM , Lucas Lucas C , Soriano Castrejón AM , García Vicente AM . Clinicopathologic and metabolic variables from 18F-FDG PET/CT in the prediction of recurrence pattern in stage I-III non-small cell lung cancer after curative surgery . Ann Nucl Med . 2025 May ; 39 ( 5 ): 476 – 505 . doi: 10.1007/s12149-025-02021-y . PMID: 39948296 . OpenUrl CrossRef PubMed 12. Dercle L , Fronheiser M , Rizvi NA , Hellmann MD , Maier S , Hayes W , Yang H , Guo P , Fojo T , Schwartz LH , Zhao B , Leung DK . Baseline Radiomic Signature to Estimate Overall Survival in Patients With NSCLC . J Thorac Oncol . 2023 May ; 18 ( 5 ): 587 – 598 . doi: 10.1016/j.jtho.2022.12.019 . PMID: 36646209 . OpenUrl CrossRef PubMed 13. Li HJ , Qiu ZB , Wang MM , Zhang C , Hong HZ , Fu R , Peng LS , Huang C , Cui Q , Zhang JT , Ren JY , Jiang L , Wu YL , Zhong WZ . Radiomics-Based Support Vector Machine Distinguishes Molecular Events Driving the Progression of Lung Adenocarcinoma . J Thorac Oncol . 2025 Jan ; 20 ( 1 ): 52 – 64 . doi: 10.1016/j.jtho.2024.09.1431 . PMID: 39306192 . OpenUrl CrossRef PubMed 14. Rossi G , Barabino E , Fedeli A , Ficarra G , Coco S , Russo A , Adamo V , Buemi F , Zullo L , Dono M , De Luca G , Longo L , Dal Bello MG , Tagliamento M , Alama A , Cittadini G , Pronzato P , Genova C . Radiomic Detection of EGFR Mutations in NSCLC . Cancer Res . 2021 Feb 1; 81 ( 3 ): 724 – 731 . doi: 10.1158/0008-5472.CAN-20-0999 . PMID: 33148663 . OpenUrl Abstract / FREE Full Text 15. ↵ Lu J , Liu X , Ji X , Jiang Y , Zuo A , Guo Z , Yang S , Peng H , Sun F , Lu D . Predicting PD-L1 status in NSCLC patients using deep learning radiomics based on CT images . Sci Rep . 2025 Apr 11; 15 ( 1 ): 12495 . doi: 10.1038/s41598-025-91575-y . PMID: 40216830 . OpenUrl CrossRef PubMed 16. ↵ Jha AK , Mithun S , Jaiswar V , Sherkhane UB , Purandare NC , Prabhash K , Rangarajan V , Dekker A , Wee L , Traverso A . Repeatability and reproducibility study of radiomic features on a phantom and human cohort . Sci Rep . 2021 Jan 21; 11 ( 1 ): 2055 . doi: 10.1038/s41598-021-81526-8 . PMID: 33479392 . OpenUrl CrossRef PubMed 17. Haarburger C , Müller-Franzes G , Weninger L , Kuhl C , Truhn D , Merhof D . Radiomics feature reproducibility under inter-rater variability in segmentations of CT images . Sci Rep . 2020 Jul 29; 10 ( 1 ): 12688 . doi: 10.1038/s41598-020-69534-6 . PMID: 32728098 . OpenUrl CrossRef PubMed 18. ↵ Rinaldi L , De Angelis SP , Raimondi S , Rizzo S , Fanciullo C , Rampinelli C , Mariani M , Lascialfari A , Cremonesi M , Orecchia R , Origgi D , Botta F . Reproducibility of radiomic features in CT images of NSCLC patients: an integrative analysis on the impact of acquisition and reconstruction parameters . Eur Radiol Exp . 2022 Jan 25; 6 ( 1 ): 2 . doi: 10.1186/s41747-021-00258-6 . PMID: 35075539 . OpenUrl CrossRef PubMed 19. ↵ Ge G , Zhang J . Feature selection methods and predictive models in CT lung cancer radiomics . J Appl Clin Med Phys . 2023 Jan ; 24 ( 1 ): e13869 . doi: 10.1002/acm2.13869 . PMID: 36527376 . OpenUrl CrossRef PubMed 20. ↵ Primakov SP , Ibrahim A , van Timmeren JE , Wu G , Keek SA , Beuque M , Granzier RWY , Lavrova E , Scrivener M , Sanduleanu S , Kayan E , Halilaj I , Lenaers A , Wu J , Monshouwer R , Geets X , Gietema HA , Hendriks LEL , Morin O , Jochems A , Woodruff HC , Lambin P . Automated detection and segmentation of non-small cell lung cancer computed tomography images . Nat Commun . 2022 Jun 14; 13 ( 1 ): 3423 . doi: 10.1038/s41467-022-30841-3 . PMID: 35701415 . OpenUrl CrossRef PubMed 21. Xie H , Chen Z , Deng J , Zhang J , Duan H , Li Q . Automatic segmentation of the gross target volume in radiotherapy for lung cancer using transresSEUnet 2.5D Network . J Transl Med . 2022 Nov 12; 20 ( 1 ): 524 . doi: 10.1186/s12967-022-03732-w . PMID: 36371220 . OpenUrl CrossRef PubMed 22. Carles M , Kuhn D , Fechter T , Baltas D , Mix M , Nestle U , Grosu AL , Martí-Bonmatí L , Radicioni G , Gkika E . Correction: 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 Dec ; 34 ( 12 ): 8039 . doi: 10.1007/s00330-024-10826-0 . PMID: 38662100 . OpenUrl CrossRef PubMed 23. ↵ Trimpl MJ , Campbell S , Panakis N , Ajzensztejn D , Burke E , Ellis S , Johnstone P , Doyle E , Towers R , Higgins G , Bernard C , Hustinx R , Vallis KA , Stride EPJ , Gooding MJ . Deep learning-assisted interactive contouring of lung cancer: Impact on contouring time and consistency . Radiother Oncol . 2024 Nov ; 200 : 110500 . doi: 10.1016/j.radonc.2024.110500 . PMID: 39236985 . OpenUrl CrossRef PubMed 24. ↵ Yan M , Zhang Z , Tian J , Yu J , Dekker A , Ruysscher D , Wee L , Zhao L . Whole lung radiomic features are associated with overall survival in patients with locally advanced non-small cell lung cancer treated with definitive radiotherapy . Radiat Oncol . 2025 Jan 17; 20 ( 1 ): 9 . doi: 10.1186/s13014-025-02583-1 . PMID: 39825409 . OpenUrl CrossRef PubMed 25. ↵ Coroller TP , Agrawal V , Huynh E , Narayan V , Lee SW , Mak RH , Aerts HJWL . Radiomic-Based Pathological Response Prediction from Primary Tumors and Lymph Nodes in NSCLC . J Thorac Oncol . 2017 Mar ; 12 ( 3 ): 467 – 476 . doi: 10.1016/j.jtho.2016.11.2226 . PMID: 27903462 . OpenUrl CrossRef PubMed 26. Hindocha S , Charlton TG , Linton-Reid K , Hunter B , Chan C , Ahmed M , Greenlay EJ , Orton M , Bunce C , Lunn J , Doran SJ , Ahmad S , McDonald F , Locke I , Power D , Blackledge M , Lee RW , Aboagye EO . Gross tumour volume radiomics for prognostication of recurrence & death following radical radiotherapy for NSCLC . NPJ Precis Oncol . 2022 Oct 27; 6 ( 1 ): 77 . doi: 10.1038/s41698-022-00322-3. Erratum in: NPJ Precis Oncol. 2022 Nov 23;6(1):87. doi: 10.1038/s41698-022-00332-1 . PMID: 36302938 . OpenUrl CrossRef PubMed 27. Vaidya P , Bera K , Gupta A , Wang X , Corredor G , Fu P , Beig N , Prasanna P , Patil P , Velu P , Rajiah P , Gilkeson R , Feldman M , Choi H , Velcheti V , Madabhushi A . CT derived radiomic score for predicting the added benefit of adjuvant chemotherapy following surgery in Stage I , II resectable Non-Small Cell Lung Cancer: a retrospective multi-cohort study for outcome prediction. Lancet Digit Health . 2020 Mar ; 2 ( 3 ): e116 – e128 . doi: 10.1016/s2589-7500(20)30002-9 . PMID: 32123864 . OpenUrl CrossRef PubMed 28. ↵ Wang T , Deng J , She Y , Zhang L , Wang B , Ren Y , Wu J , Xie D , Sun X , Chen C . Radiomics Signature Predicts the Recurrence-Free Survival in Stage I Non-Small Cell Lung Cancer . Ann Thorac Surg . 2020 Jun ; 109 ( 6 ): 1741 – 1749 . doi: 10.1016/j.athoracsur.2020.01.010 . PMID: 32087134 . OpenUrl CrossRef PubMed 29. ↵ Pignon JP , Tribodet H , Scagliotti GV , Douillard JY , Shepherd FA , Stephens RJ , Dunant A , Torri V , Rosell R , Seymour L , Spiro SG , Rolland E , Fossati R , Aubert D , Ding K , Waller D , Le Chevalier T; LACE Collaborative Group. Lung adjuvant cisplatin evaluation: a pooled analysis by the LACE Collaborative Group . J Clin Oncol . 2008 Jul 20; 26 ( 21 ): 3552 – 9 . doi: 10.1200/JCO.2007.13.9030 . PMID: 18506026 . OpenUrl Abstract / FREE Full Text 30. ↵ Forde PM , Spicer J , Lu S , Provencio M , Mitsudomi T , Awad MM , Felip E , Broderick SR , Brahmer JR , Swanson SJ , Kerr K , Wang C , Ciuleanu TE , Saylors GB , Tanaka F , Ito H , Chen KN , Liberman M , Vokes EE , Taube JM , Dorange C , Cai J , Fiore J , Jarkowski A , Balli D , Sausen M , Pandya D , Calvet CY , Girard N; CheckMate 816 Investigators. Neoadjuvant Nivolumab plus Chemotherapy in Resectable Lung Cancer . N Engl J Med . 2022 May 26; 386 ( 21 ): 1973 – 1985 . doi: 10.1056/NEJMoa2202170 . PMID: 35403841 . OpenUrl CrossRef PubMed 31. ↵ Spicer JD , Garassino MC , Wakelee H , Liberman M , Kato T , Tsuboi M , Lee SH , Chen KN , Dooms C , Majem M , Eigendorff E , Martinengo GL , Bylicki O , Rodríguez-Abreu D , Chaft JE , Novello S , Yang J , Arunachalam A , Keller SM , Samkari A , Gao S ; KEYNOTE-671 Investigators. Neoadjuvant pembrolizumab plus chemotherapy followed by adjuvant pembrolizumab compared with neoadjuvant chemotherapy alone in patients with early-stage non-small-cell lung cancer (KEYNOTE-671): a randomised, double-blind, placebo-controlled, phase 3 trial . Lancet . 2024 Sep 28; 404 ( 10459 ):1240-1252. doi: 10.1016/S0140-6736(24)01756-2 . PMID: 39288781 . OpenUrl CrossRef PubMed 32. ↵ Postmus PE , Kerr KM , Oudkerk M , Senan S , Waller DA , Vansteenkiste J , Escriu C , Peters S; ESMO Guidelines Committee. Early and locally advanced non-small-cell lung cancer (NSCLC): ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up . Ann Oncol . 2017 Jul 1; 28 (suppl_4):iv1-iv21. doi: 10.1093/annonc/mdx222 . PMID: 28881918 . OpenUrl CrossRef PubMed 33. ↵ Moss C , Haire A , Cahill F , Enting D , Hughes S , Smith D , Sawyer E , Davies A , Zylstra J , Haire K , Rigg A , Van Hemelrijck M . Guy’s cancer cohort - real world evidence for cancer pathways . BMC Cancer . 2020 Mar 17; 20 ( 1 ): 187 . doi: 10.1186/s12885-020-6667-0 . PMID: 32178645 . OpenUrl CrossRef PubMed 34. ↵ Isensee F , Jaeger PF , Kohl SAA , Petersen J , Maier-Hein KH . nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation . Nat Methods . 2021 Feb ; 18 ( 2 ): 203 – 211 . doi: 10.1038/s41592-020-01008-z . PMID: 33288961 . OpenUrl CrossRef PubMed 35. ↵ Bhamani A , Creamer A , Verghese P , Prendecki R , Horst C , Tisi S , Hall H , Khaw CR , Mullin M , McCabe J , Gyertson K , Bowyer V , Arancon D , Eng J , Bojang F , Levermore C , Hacker AM , Arthur-Darkwa E , Farrelly L , Patel A , Lock S , Shaw A , Banka R , Bhowmik A , Ekeowa U , Mangera Z , Valerio C , Ricketts WM , Mohammed A , O’Shaughnessy T , Navani N , Quaife SL , Nair A , Devaraj A; SUMMIT consortium; Dickson JL, Hackshaw A , Janes SM. Low-dose CT for lung cancer screening in a high-risk population (SUMMIT): a prospective, longitudinal cohort study. Lancet Oncol . 2025 May ; 26 ( 5 ): 609 – 619 . doi: 10.1016/S1470-2045(25)00082-8 . PMID: 40154514 . OpenUrl CrossRef PubMed 36. ↵ Zhao B , Schwartz LH , Kris MG , Riely GJ . Coffee-break lung CT collection with scan images reconstructed at multiple imaging parameters (Version 3) [Dataset] . The Cancer Imaging Archive ; 2015 . Available from : doi: 10.7937/k9/tcia.2015.u1x8a5nr . Accessed on 2024 Oct. OpenUrl CrossRef 37. Wee L , Aerts HJL , Kalendralis P , Dekker A . Data from NSCLC-Radiomics-Interobserver1 [Data set] . The Cancer Imaging Archive ; 2019 . Available from : doi: 10.7937/tcia.2019.cwvlpd26 _Accessed on 2024 Oct. OpenUrl CrossRef 38. Aerts HJWL , Wee L , Rios Velazquez E , Leijenaar RTH , Parmar C , Grossmann P , Carvalho S , Bussink J , Monshouwer R , Haibe-Kains B , Rietveld D , Hoebers F , Rietbergen MM , Leemans CR , Dekker A , Quackenbush J , Gillies RJ , Lambin P . Data From NSCLC-Radiomics (version 4) [Data set] . The Cancer Imaging Archive ; 2014 . Available from : doi: 10.7937/K9/TCIA.2015.PF0M9REI . Accessed on 2024 Oct. OpenUrl CrossRef 39. Bakr S , Gevaert O , Echegaray S , Ayers K , Zhou M , Shafiq M , Zheng H , Zhang W , Leung A , Kadoch M , Shrager J , Quon A , Rubin D , Plevritis S , Napel S . Data for NSCLC Radiogenomics (Version 4) [Data set] . The Cancer Imaging Archive ; 2017 . Available from : doi: 10.7937/K9/TCIA.2017.7hs46erv . Accessed on 2024 Oct. OpenUrl CrossRef 40. ↵ Goldgof D. , Hall L. , Hawkins S.H. , Schabath M.B. , Stringfield O. , Garcia A. , Balagurunathan Y. , Kim J. , Eschrich S. , Berglund A.E. , Gatenby R. , Gillies RJ . Long and Short Survival in Adenocarcinoma Lung CTs [Data set] . The Cancer Imaging Archive ; 2017 . Available from : doi: 10.7937/K9/TCIA.2017.0tv7b9x1 . Accessed on 2024 Oct. OpenUrl CrossRef 41. ↵ Wang K , Wang Y , Zhan B , Yang Y , Zu C , Wu X , Zhou J , Nie D , Zhou L . An Efficient Semi-Supervised Framework with Multi-Task and Curriculum Learning for Medical Image Segmentation . Int J Neural Syst . 2022 Sep ; 32 ( 9 ): 2250043 . doi: 10.1142/S0129065722500435 . PMID: 35912583 . OpenUrl CrossRef PubMed 42. ↵ Liu L , Wolterink JM , Brune C , Veldhuis RNJ . Anatomy-aided deep learning for medical image segmentation: a review . Phys Med Biol . 2021 May 26; 66 ( 11 ). doi: 10.1088/1361-6560/abfbf4 . PMID: 33906186 . OpenUrl CrossRef PubMed 43. ↵ Wasserthal J , Breit HC , Meyer MT , Pradella M , Hinck D , Sauter AW , Heye T , Boll DT , Cyriac J , Yang S , Bach M , Segeroth M . TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images . Radiol Artif Intell . 2023 Jul 5; 5 ( 5 ): e230024 . doi: 10.1148/ryai.230024 . PMID: 37795137 . OpenUrl CrossRef PubMed 44. ↵ Wu L , Lou X , Kong N , Xu M , Gao C . Can quantitative peritumoral CT radiomics features predict the prognosis of patients with non-small cell lung cancer? A systematic review . Eur Radiol . 2023 Mar ; 33 ( 3 ): 2105 – 2117 . doi: 10.1007/s00330-022-09174-8 . PMID: 36307554 . OpenUrl CrossRef PubMed 45. ↵ Teng X , Zhang J , Zwanenburg A , Sun J , Huang Y , Lam S , Zhang Y , Li B , Zhou T , Xiao H , Liu C , Li W , Han X , Ma Z , Li T , Cai J . Building reliable radiomic models using image perturbation . Sci Rep . 2022 Jun 16; 12 ( 1 ): 10035 . doi: 10.1038/s41598-022-14178-x . PMID: 35710850 . OpenUrl CrossRef PubMed 46. ↵ van Griethuysen JJM , Fedorov A , Parmar C , Hosny A , Aucoin N , Narayan V , Beets-Tan RGH , Fillion-Robin JC , Pieper S , Aerts HJWL . Computational Radiomics System to Decode the Radiographic Phenotype . Cancer Res . 2017 Nov 1; 77 ( 21 ): e104 – e107 . doi: 10.1158/0008-5472.CAN-17-0339 . PMID: 29092951 . OpenUrl Abstract / FREE Full Text 47. ↵ Riley RD , Ensor J , Snell KIE , Harrell FE Jr . , Martin GP , Reitsma JB , Moons KGM , Collins G , van Smeden M . Calculating the sample size required for developing a clinical prediction model . BMJ . 2020 Mar 18; 368 : m441 . doi: 10.1136/bmj.m441 . PMID: 32188600 . OpenUrl FREE Full Text 48. ↵ Strobl C , Boulesteix AL , Zeileis A , Hothorn T . Bias in random forest variable importance measures: illustrations, sources and a solution . BMC Bioinformatics . 2007 Jan 25; 8 : 25 . doi: 10.1186/1471-2105-8-25 . PMID: 17254353 . OpenUrl CrossRef PubMed 49. ↵ Vaassen F , Hazelaar C , Vaniqui A , Gooding M , van der Heyden B , Canters R , van Elmpt W . Evaluation of measures for assessing time-saving of automatic organ-at-risk segmentation in radiotherapy . Phys Imaging Radiat Oncol . 2019 Dec 17; 13 : 1 – 6 . doi: 10.1016/j.phro.2019.12.001 . PMID: 33458300 . OpenUrl CrossRef PubMed 50. ↵ Mercieca S , Belderbos JSA , van Herk M . Challenges in the target volume definition of lung cancer radiotherapy . Transl Lung Cancer Res . 2021 Apr ; 10 ( 4 ): 1983 – 1998 . doi: 10.21037/tlcr-20-627 . PMID: 34012808 . OpenUrl CrossRef PubMed 51. ↵ Isler IS , Mohaisen D , Lisle C , Turgut D , Bagci U. Uncertainty-guided coarse-to-fine tumor segmentation with anatomy-aware post-processing . arXiv [Preprint] . 2025 . Available from: https://arxiv.org/abs/2504.12215 . Accessed on 2025 May . 52. ↵ Mathai TS , Liu B , Summers RM . Segmentation of mediastinal lymph nodes in CT with anatomical priors . Int J Comput Assist Radiol Surg . 2024 Aug ; 19 ( 8 ): 1537 – 1544 . doi: 10.1007/s11548-024-03165-4 . PMID: 38740719 . OpenUrl CrossRef PubMed 53. ↵ Myronenko A , Hatamizadeh A. 3D kidneys and kidney tumor semantic segmentation using boundary-aware networks . arXiv [Preprint] . 2019 . Available from: https://arxiv.org/abs/1909.06684 . Accessed on 2025 May . 54. ↵ Wang X , Li X , Du R , Zhong Y , Lu Y , Song T . Anatomical Prior-Based Automatic Segmentation for Cardiac Substructures from Computed Tomography Images . Bioengineering (Basel ). 2023 Oct 31; 10 ( 11 ): 1267 . doi: 10.3390/bioengineering10111267 . PMID: 38002391 . OpenUrl CrossRef PubMed 55. Himmetoglu M , Ciernik I , Konukoglu E. Learning to segment anatomy and lesions from disparately labeled sources in brain MRI . arXiv [Preprint] . 2025 . Available from: https://arxiv.org/abs/2503.18840 56. Sital C , Brosch T , Tio D , Raaijmakers AJ , Weese J. 3D medical image segmentation with labeled and unlabeled data using autoencoders at the example of liver segmentation in CT images . arXiv [Preprint] . 2020 . Available from: https://arxiv.org/abs/2003.07923 . Accessed on 2025 May . 57. ↵ Prasad AV , Mathai T , Mukherjee P , Liu J , Summers RM . Leveraging anatomical priors for automated pancreas segmentation on abdominal CT . In: Medical Imaging 2025 : Computer-Aided Diagnosis. SPIE; 2025. Paper 12926-54. doi: 10.1117/12.3004040 . OpenUrl CrossRef 58. ↵ Mathai T , Hou B , Summers RM . Longitudinal assessment of lung lesion burden in CT . In: Medical Imaging 2025 : Computer-Aided Diagnosis. SPIE; 2025. Paper 12926-57. doi: 10.1117/12.3004065 . OpenUrl CrossRef 59. ↵ Asha V , Bhavanishankar K. Advanced lung nodule segmentation and classification for early detection of lung cancer using SAM and transfer learning . arXiv [Preprint] . 2024 . Available from: https://arxiv.org/abs/2501.00586 . Accessed on 2025 May . 60. ↵ Denzler S , Vuong D , Bogowicz M , Pavic M , Frauenfelder T , Thierstein S , Eboulet EI , Maurer B , Schniering J , Gabryś HS , Schmitt-Opitz I , Pless M , Foerster R , Guckenberger M , Tanadini-Lang S . Impact of CT convolution kernel on robustness of radiomic features for different lung diseases and tissue types . Br J Radiol . 2021 Apr 1; 94 ( 1120 ):20200947. doi: 10.1259/bjr.20200947 . PMID: 33544646 . OpenUrl CrossRef PubMed 61. ↵ Reiazi R , Abbas E , Famiyeh P , Rezaie A , Kwan JYY , Patel T , Bratman SV , Tadic T , Liu FF , Haibe-Kains B . The impact of the variation of imaging parameters on the robustness of Computed Tomography radiomic features: A review . Comput Biol Med . 2021 Jun ; 133 : 104400 . doi: 10.1016/j.compbiomed.2021.104400 . PMID: 33930766 . OpenUrl CrossRef PubMed 62. ↵ Hu F , Chen AA , Horng H , Bashyam V , Davatzikos C , Alexander-Bloch A , Li M , Shou H , Satterthwaite TD , Yu M , Shinohara RT . Image harmonization: A review of statistical and deep learning methods for removing batch effects and evaluation metrics for effective harmonization . Neuroimage . 2023 Jul 1; 274 : 120125 . doi: 10.1016/j.neuroimage.2023.120125 . PMID: 37084926 . OpenUrl CrossRef PubMed 63. ↵ Wu J , Li C , Gensheimer M , Padda S , Kato F , Shirato H , Wei Y , Schönlieb CB , Price SJ , Jaffray D , Heymach J , Neal JW , Loo BW Jr , Wakelee H , Diehn M , Li R . Radiological tumor classification across imaging modality and histology . Nat Mach Intell . 2021 Sep ; 3 : 787 – 798 . doi: 10.1038/s42256-021-00377-0 . PMID: 34841195 . OpenUrl CrossRef PubMed 64. ↵ Braman N , Prasanna P , Bera K , Alilou M , Khorrami M , Leo P , Etesami M , Vulchi M , Turk P , Gupta A , Jain P , Fu P , Pennell N , Velcheti V , Abraham J , Plecha D , Madabhushi A . Novel Radiomic Measurements of Tumor-Associated Vasculature Morphology on Clinical Imaging as a Biomarker of Treatment Response in Multiple Cancers . Clin Cancer Res . 2022 Oct 14; 28 ( 20 ): 4410 – 4424 . doi: 10.1158/1078-0432.CCR-21-4148 . PMID: 35727603 . OpenUrl CrossRef PubMed 65. ↵ Hosny A , Parmar C , Coroller TP , Grossmann P , Zeleznik R , Kumar A , Bussink J , Gillies RJ , Mak RH , Aerts HJWL . Deep learning for lung cancer prognostication: A retrospective multi-cohort radiomics study . PLoS Med . 2018 Nov 30; 15 ( 11 ): e1002711 . doi: 10.1371/journal.pmed.1002711 . PMID: 30500819 . OpenUrl CrossRef PubMed 66. ↵ Orlhac F , Eertink JJ , Cottereau AS , Zijlstra JM , Thieblemont C , Meignan M , Boellaard R , Buvat I . A Guide to ComBat Harmonization of Imaging Biomarkers in Multicenter Studies . J Nucl Med . 2022 Feb ; 63 ( 2 ): 172 – 179 . doi: 10.2967/jnumed.121.262464 . PMID: 34531263 . OpenUrl Abstract / FREE Full Text View the discussion thread. Back to top Previous Next Posted June 18, 2025. Download PDF Supplementary Material Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following A Deep Learning Lung Cancer Segmentation Pipeline to Facilitate CT-based Radiomics Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share A Deep Learning Lung Cancer Segmentation Pipeline to Facilitate CT-based Radiomics Alfred Chung Pui So , Daryl Cheng , Shahab Aslani , Mehran Azimbagirad , Daisuke Yamada , Roberta Dunn , Eleni Josephides , Ellie McDowall , Annie-Rose Henry , Andrea Bille , Nishanth Sivarasan , Adam Pennycuick , Eleni Karapanagiotou , Joseph Jacob medRxiv 2025.06.17.25329213; doi: https://doi.org/10.1101/2025.06.17.25329213 Share This Article: Copy Citation Tools A Deep Learning Lung Cancer Segmentation Pipeline to Facilitate CT-based Radiomics Alfred Chung Pui So , Daryl Cheng , Shahab Aslani , Mehran Azimbagirad , Daisuke Yamada , Roberta Dunn , Eleni Josephides , Ellie McDowall , Annie-Rose Henry , Andrea Bille , Nishanth Sivarasan , Adam Pennycuick , Eleni Karapanagiotou , Joseph Jacob medRxiv 2025.06.17.25329213; doi: https://doi.org/10.1101/2025.06.17.25329213 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Radiology and Imaging Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (299) Cardiovascular Medicine (4425) Dentistry and Oral Medicine (443) Dermatology (382) Emergency Medicine (607) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1507) Epidemiology (15221) Forensic Medicine (30) Gastroenterology (1123) Genetic and Genomic Medicine (6588) Geriatric Medicine (667) Health Economics (997) Health Informatics (4524) Health Policy (1368) Health Systems and Quality Improvement (1612) Hematology (540) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15910) Intensive Care and Critical Care Medicine (1103) Medical Education (623) Medical Ethics (145) Nephrology (667) Neurology (6588) Nursing (346) Nutrition (998) Obstetrics and Gynecology (1143) Occupational and Environmental Health (956) Oncology (3331) Ophthalmology (970) Orthopedics (369) Otolaryngology (420) Pain Medicine (435) Palliative Medicine (129) Pathology (663) Pediatrics (1690) Pharmacology and Therapeutics (691) Primary Care Research (710) Psychiatry and Clinical Psychology (5440) Public and Global Health (9219) Radiology and Imaging (2195) Rehabilitation Medicine and Physical Therapy (1369) Respiratory Medicine (1196) Rheumatology (593) Sexual and Reproductive Health (710) Sports Medicine (529) Surgery (710) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ffacde39e2ac13d',t:'MTc3OTQ0MjI0OA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();