Pub Date : 2026-01-01DOI: 10.1016/j.acra.2025.09.037
Chunxiao Wang , Yuxin Li , Yang Ji, Kang Yu, Chunhui Qin, Ling Liu, Yunjia Shuai, Jiahui Chen, Ao Li, Tong Zhang
Background
Determining predictive biomarkers for immunotherapy response in non-small cell lung cancer (NSCLC) patients is a complex task.
Objective
This research aimed to develop a multimodal model (CRDL) integrating clinical data, deep learning (DL), and radiomics (Rad) to predict immune responses in NSCLC patients receiving checkpoint blockade therapies. This study also evaluated whether CRDL outperforms unimodal, pre-fusion models (Pre-FMs) and post-fusion models (Post-FMs).
Methods
This is a retrospective study that utilized data from 228 lung cancer patients at the Memorial Sloan Kettering Cancer Center, with varying Programmed Death-Ligand 1(PD-L1) expression levels among the patients. 228 NSCLC patients were randomly divided into two groups in a 7:3 ratio: the training cohort (n = 159) and the validation cohort (n = 69). Image histological features were extracted using the "PyRadiomics" package, and DL features were obtained through the deep convolutional neural network from chest computed tomography images, and clinical data from the patients were also collected. Feature reduction was performed using t-tests and the Least absolute shrinkage and selection operator regression. Unimodal modal and Pre-FMs were constructed using random forests, while the post-fusion model was developed using a support vector machine approach. The performance of the model was measured by the area under the receiver operating characteristic curve (AUC).
Results
512 DL features and 382 Rad features were extracted. The CRDL model demonstrated superior performance with AUC values of 0.884 in the validation dataset and 0.976 in the training dataset, surpassing the best DL model in both unimodal and pre-fusion settings, which had training and validation AUCs of 0.854 and 0.749.
Conclusion
The CRDL model accurately forecasts immunotherapy responses in NSCLC patients, offering one dependable non-invasive test.
{"title":"Predictions of Response in Non-small Cell Lung Cancer Patients Treated with Immune Checkpoint Inhibitors Using Clinical Data, Deep Learning, and Radiomics","authors":"Chunxiao Wang , Yuxin Li , Yang Ji, Kang Yu, Chunhui Qin, Ling Liu, Yunjia Shuai, Jiahui Chen, Ao Li, Tong Zhang","doi":"10.1016/j.acra.2025.09.037","DOIUrl":"10.1016/j.acra.2025.09.037","url":null,"abstract":"<div><h3>Background</h3><div>Determining predictive biomarkers for immunotherapy response in non-small cell lung cancer (NSCLC) patients is a complex task.</div></div><div><h3>Objective</h3><div>This research aimed to develop a multimodal model (CRDL) integrating clinical data, deep learning (DL), and radiomics (Rad) to predict immune responses in NSCLC patients receiving checkpoint blockade therapies. This study also evaluated whether CRDL outperforms unimodal, pre-fusion models (Pre-FMs) and post-fusion models (Post-FMs).</div></div><div><h3>Methods</h3><div>This is a retrospective study that utilized data from 228 lung cancer patients at the Memorial Sloan Kettering Cancer Center, with varying Programmed Death-Ligand 1(PD-L1) expression levels among the patients. 228 NSCLC patients were randomly divided into two groups in a 7:3 ratio: the training cohort (<em>n<!--> </em>=<!--> <!-->159) and the validation cohort (<em>n<!--> </em>=<!--> <!-->69). Image histological features were extracted using the \"PyRadiomics\" package, and DL features were obtained through the deep convolutional neural network from chest computed tomography images, and clinical data from the patients were also collected. Feature reduction was performed using t-tests and the Least absolute shrinkage and selection operator regression. Unimodal modal and Pre-FMs were constructed using random forests, while the post-fusion model was developed using a support vector machine approach. The performance of the model was measured by the area under the receiver operating characteristic curve (AUC).</div></div><div><h3>Results</h3><div>512 DL features and 382 Rad features were extracted. The CRDL model demonstrated superior performance with AUC values of 0.884 in the validation dataset and 0.976 in the training dataset, surpassing the best DL model in both unimodal and pre-fusion settings, which had training and validation AUCs of 0.854 and 0.749.</div></div><div><h3>Conclusion</h3><div>The CRDL model accurately forecasts immunotherapy responses in NSCLC patients, offering one dependable non-invasive test.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 1","pages":"Pages 236-254"},"PeriodicalIF":3.9,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145294394","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-01DOI: 10.1016/j.acra.2025.10.006
Fukun Shi , Xianqing Ren , Qian Xu , Jiameng Si , Yihao Yan , Junjie Shu , Shengli Shi , Ke Jin , Fenfen Li , Jiajia Zhang , Lan Zhang
Rationale and Objectives
To develop and validate an interpretable radiomics model based on pituitary MRI to predict growth hormone deficiency (GHD) in children with short stature.
Methods
This retrospective multicenter study enrolled 202 children (105 GHD, 97 idiopathic short stature [ISS]) as an internal cohort (7:3 ratio for training/testing cohorts) from institution I, and 138 children (61 GHD, 77 ISS) from institution II and institution III as an external validation cohort. Radiomics features were selected by SelectKBest and least absolute shrinkage and selection operator (LASSO), subsequently used to construct six machine learning models. Diagnostic performance of model was evaluated by the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and calibration curves. The interpretability of the model was assessed using Shapley additive explanations (SHAP).
Results
A total of 17 radiomics features were selected. Among all classifiers, support vector machine (SVM)-based radiomics model exhibited the highest diagnostic performance, with AUCs of 0.877 (95% CI: 0.813, 0.928), 0.878 (95% CI: 0.786, 0.951), and 0.885 (95% CI: 0.833, 0.937) in training, testing, and external validation cohorts, respectively. The SVM-integrated clinical-radiomics model yielded comparable efficacy, with AUCs of 0.874 (95% CI: 0.812, 0.928), 0.878 (95% CI: 0.786, 0.952), and 0.889 (95% CI: 0.830, 0.939) across the same cohorts. Both radiomics-based models significantly outperformed the clinical model (all p<0.001), while no statistically significant difference was observed between the radiomics and clinical-radiomics models (all p>0.05). The SHAP analysis identified three key radiomics features with significant differences between GHD and ISS groups (all p<0.001).
Conclusions
The interpretable radiomics-driven SVM model effectively predicts GH levels, providing a clinically viable, non-invasive alternative to GH stimulation test in children with short stature.
{"title":"An Interpretable Radiomics Model Based on Pituitary MRI to Predict Growth Hormone Deficiency in Short-statured Children: A Multicenter Study","authors":"Fukun Shi , Xianqing Ren , Qian Xu , Jiameng Si , Yihao Yan , Junjie Shu , Shengli Shi , Ke Jin , Fenfen Li , Jiajia Zhang , Lan Zhang","doi":"10.1016/j.acra.2025.10.006","DOIUrl":"10.1016/j.acra.2025.10.006","url":null,"abstract":"<div><h3>Rationale and Objectives</h3><div>To develop and validate an interpretable radiomics model based on pituitary MRI to predict growth hormone deficiency (GHD) in children with short stature.</div></div><div><h3>Methods</h3><div>This retrospective multicenter study enrolled 202 children (105 GHD, 97 idiopathic short stature [ISS]) as an internal cohort (7:3 ratio for training/testing cohorts) from institution I, and 138 children (61 GHD, 77 ISS) from institution II and institution III as an external validation cohort. Radiomics features were selected by SelectKBest and least absolute shrinkage and selection operator (LASSO), subsequently used to construct six machine learning models. Diagnostic performance of model was evaluated by the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and calibration curves. The interpretability of the model was assessed using Shapley additive explanations (SHAP).</div></div><div><h3>Results</h3><div>A total of 17 radiomics features were selected. Among all classifiers, support vector machine (SVM)-based radiomics model exhibited the highest diagnostic performance, with AUCs of 0.877 (95% <em>CI</em>: 0.813, 0.928), 0.878 (95% <em>CI</em>: 0.786, 0.951), and 0.885 (95% <em>CI</em>: 0.833, 0.937) in training, testing, and external validation cohorts, respectively. The SVM-integrated clinical-radiomics model yielded comparable efficacy, with AUCs of 0.874 (95% <em>CI</em>: 0.812, 0.928), 0.878 (95% <em>CI</em>: 0.786, 0.952), and 0.889 (95% <em>CI</em>: 0.830, 0.939) across the same cohorts. Both radiomics-based models significantly outperformed the clinical model (all <em>p</em><0.001), while no statistically significant difference was observed between the radiomics and clinical-radiomics models (all <em>p</em>>0.05). The SHAP analysis identified three key radiomics features with significant differences between GHD and ISS groups (all <em>p</em><0.001).</div></div><div><h3>Conclusions</h3><div>The interpretable radiomics-driven SVM model effectively predicts GH levels, providing a clinically viable, non-invasive alternative to GH stimulation test in children with short stature.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 1","pages":"Pages 168-179"},"PeriodicalIF":3.9,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145395046","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-01DOI: 10.1016/j.acra.2025.09.039
Kelly M. Gillen PhD, MBA , Mert Şişman , Alan Wu , Arindam RoyChoudhury PhD , Ajay Gupta MD, MS
Rationale and Objectives
With a budget of almost $48 billion in 2024, including $37 billion allocated for extramural funding, the National Institutes of Health (NIH) is the major funding source for biomedical research in the United States. Given the multi-faceted impact of NIH funding on academic institutions and their communities, we sought to characterize trends in research project grant funding to departments of radiology.
This study aimed to assess trends in R01 and R01 equivalent award funding to departments of radiology from 2014 to 2024.
Materials and Methods
All funding data were retrieved from NIH RePORTER (Research Portfolio Online Reporting Tools) and limited to R01 and R01 equivalent awards (R01+) to all clinical departments (ACDs) from NIH FY 2014 and FY 2024. Awards and funding data included ACDs as categorized by the Blue Ridge Institute for Medical Research. Information on principal investigator (PI) advanced degrees was obtained by web searches and visiting the PI’s faculty page through their respective academic institution.
Results
From 2014 to 2024, there was a 54.3% increase in the number of R01s awarded to radiology as compared to a 31.7% increase in R01s awarded to ACDs. There was a 69.0% increase in the number of R01+s awarded to radiology as compared to a 34.4% increase in R01+s awarded to ACDs during this same period.
Conclusion
Since FY2014, there has been an increase in funding from the NIH to ACDs and specifically to radiology, but departments of radiology are outpacing ACDs in several key R01 and R01+ funding metrics, including greater increases in the number of awards.
{"title":"Trends in R01 and R01 Equivalent Funding to Radiology","authors":"Kelly M. Gillen PhD, MBA , Mert Şişman , Alan Wu , Arindam RoyChoudhury PhD , Ajay Gupta MD, MS","doi":"10.1016/j.acra.2025.09.039","DOIUrl":"10.1016/j.acra.2025.09.039","url":null,"abstract":"<div><h3>Rationale and Objectives</h3><div>With a budget of almost $48 billion in 2024, including $37 billion allocated for extramural funding, the National Institutes of Health (NIH) is the major funding source for biomedical research in the United States. Given the multi-faceted impact of NIH funding on academic institutions and their communities, we sought to characterize trends in research project grant funding to departments of radiology.</div><div>This study aimed to assess trends in R01 and R01 equivalent award funding to departments of radiology from 2014 to 2024.</div></div><div><h3>Materials and Methods</h3><div>All funding data were retrieved from NIH RePORTER (Research Portfolio Online Reporting Tools) and limited to R01 and R01 equivalent awards (R01<sup>+</sup>) to all clinical departments (ACDs) from NIH FY 2014 and FY 2024. Awards and funding data included ACDs as categorized by the Blue Ridge Institute for Medical Research. Information on principal investigator (PI) advanced degrees was obtained by web searches and visiting the PI’s faculty page through their respective academic institution.</div></div><div><h3>Results</h3><div>From 2014 to 2024, there was a 54.3% increase in the number of R01s awarded to radiology as compared to a 31.7% increase in R01s awarded to ACDs. There was a 69.0% increase in the number of R01<sup>+</sup>s awarded to radiology as compared to a 34.4% increase in R01<sup>+</sup>s awarded to ACDs during this same period.</div></div><div><h3>Conclusion</h3><div>Since FY2014, there has been an increase in funding from the NIH to ACDs and specifically to radiology, but departments of radiology are outpacing ACDs in several key R01 and R01<sup>+</sup> funding metrics, including greater increases in the number of awards.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 1","pages":"Pages 15-19"},"PeriodicalIF":3.9,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145304239","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-01DOI: 10.1016/j.acra.2025.09.042
Deniz Esin Tekcan Sanli , Ahmet Necati Sanli
{"title":"Challenges in Imaging-Based Nomogram Models for ALNM in TNBC: Commentary on Mammography and Ultrasound Approaches","authors":"Deniz Esin Tekcan Sanli , Ahmet Necati Sanli","doi":"10.1016/j.acra.2025.09.042","DOIUrl":"10.1016/j.acra.2025.09.042","url":null,"abstract":"","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 1","pages":"Pages 82-83"},"PeriodicalIF":3.9,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145304253","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-01DOI: 10.1016/j.acra.2025.06.032
Richard Ho , Omer A. Awan MD MPH CIIP
{"title":"Navigating Entrepreneurship In and Out of the Radiology Realm","authors":"Richard Ho , Omer A. Awan MD MPH CIIP","doi":"10.1016/j.acra.2025.06.032","DOIUrl":"10.1016/j.acra.2025.06.032","url":null,"abstract":"","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 1","pages":"Pages 20-21"},"PeriodicalIF":3.9,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144644093","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-01DOI: 10.1016/j.acra.2025.09.043
Vanessa F. Schmidt MD , Jil Grethen MD , Mariya Pryadko DMD , Hannah E. Gildein MD , Matthias P. Fabritius MD , Dominik Nörenberg MD , Philipp M. Kazmierczak MD , Clemens C. Cyran MD , Rajasekhara Ayyagari MD , Sinan Deniz MD , Moritz Wildgruber MD, PhD , Max Seidensticker MD , Maciej Pech MD , Jens Ricke MD , Daniel Puhr-Westerheide MD
Rationale and Objectives
To evaluate technical, clinical, and imaging success following prostatic artery embolization (PAE) and imaging-derived predictors of treatment success based on pre- and periprocedural imaging modalities.
Material and Methods
The prospective single-center PROEMBO trial included 119 patients with symptomatic benign prostatic hyperplasia (BPH) undergoing PAE using non-spherical polyvinyl alcohol (PVA) particles. Pre- and periprocedural imaging included CT angiography (CTA), digital subtraction angiography (DSA), and cone-beam CT (CBCT). Technical success, clinical improvement (IPSS, QoL, ICIQ-SF, IIEF-EF), and imaging success (prostate volume reduction) were assessed at baseline and 1-, 6-, and 12 months of follow-ups. Imaging-derived predictors included prostatic artery anatomy, atherosclerotic plaque burden, iliac artery tortuosity, and intraprostatic density measurements after contrast administration.
Results
Median prostate volume was 66 mL (IQR, 49.5–96.5 mL). Bilateral embolization was achieved in 96/119 (80.7%), unilateral embolization in 18/119 (15.1%), and bilateral failure occurred in 5/119 patients (4.2%). IPSS improved by median of 7 points (IQR, 2–10) at 1 month, 7 points (IQR, 2.5–13) at 6 months, and 5 points (IQR, 2–12) at 12 months. MRI-based prostate volume significantly decreased by median of 6 mL (IQR, 0–16 mL), 7.5 mL (IQR, −0.75–20 mL), and 3 mL (IQR, −2–17 mL). Technical failure was significantly associated with number (p = 0.003) and cumulative area (p = 0.002) of calcified lesions along vascular access route. Iliac artery tortuosity differed significantly between groups, with higher angulation observed in patients with technical failure, e.g., bilateral angle sum of iliac bifurcation: 67.6 ± 21.2 vs. 87.8 ± 24.3 (p<0.001). PA anatomical variants, dominance patterns, and CTA-/CBCT-based density measurements showed no significant correlation with technical outcome (all p>0.05).
Conclusion
Pre- and periprocedural imaging offers valuable predictors of technical success in PAE. Atherosclerotic plaque burden and iliac artery tortuosity serve as practical markers for technical failure, thereby aiding patient selection and procedural planning.
{"title":"Pre- and Periprocedural Imaging Predicts Technical but not Clinical Success of Prostatic Artery Embolization Using Non-spherical PVA Particles—Insights From the Prospective PROEMBO Trial","authors":"Vanessa F. Schmidt MD , Jil Grethen MD , Mariya Pryadko DMD , Hannah E. Gildein MD , Matthias P. Fabritius MD , Dominik Nörenberg MD , Philipp M. Kazmierczak MD , Clemens C. Cyran MD , Rajasekhara Ayyagari MD , Sinan Deniz MD , Moritz Wildgruber MD, PhD , Max Seidensticker MD , Maciej Pech MD , Jens Ricke MD , Daniel Puhr-Westerheide MD","doi":"10.1016/j.acra.2025.09.043","DOIUrl":"10.1016/j.acra.2025.09.043","url":null,"abstract":"<div><h3>Rationale and Objectives</h3><div>To evaluate technical, clinical, and imaging success following prostatic artery embolization (PAE) and imaging-derived predictors of treatment success based on pre- and periprocedural imaging modalities.</div></div><div><h3>Material and Methods</h3><div>The prospective single-center PROEMBO trial included 119 patients with symptomatic benign prostatic hyperplasia (BPH) undergoing PAE using non-spherical polyvinyl alcohol (PVA) particles. Pre- and periprocedural imaging included CT angiography (CTA), digital subtraction angiography (DSA), and cone-beam CT (CBCT). Technical success, clinical improvement (IPSS, QoL, ICIQ-SF, IIEF-EF), and imaging success (prostate volume reduction) were assessed at baseline and 1-, 6-, and 12 months of follow-ups. Imaging-derived predictors included prostatic artery anatomy, atherosclerotic plaque burden, iliac artery tortuosity, and intraprostatic density measurements after contrast administration.</div></div><div><h3>Results</h3><div>Median prostate volume was 66 mL (IQR, 49.5–96.5<!--> <!-->mL). Bilateral embolization was achieved in 96/119 (80.7%), unilateral embolization in 18/119 (15.1%), and bilateral failure occurred in 5/119 patients (4.2%). IPSS improved by median of 7 points (IQR, 2–10) at 1 month, 7 points (IQR, 2.5–13) at 6 months, and 5 points (IQR, 2–12) at 12 months. MRI-based prostate volume significantly decreased by median of 6 mL (IQR, 0–16 mL), 7.5 mL (IQR, −0.75–20<!--> <!-->mL), and 3 mL (IQR, −2–17<!--> <!-->mL). Technical failure was significantly associated with number (p<!--> <!-->=<!--> <!-->0.003) and cumulative area (p<!--> <!-->=<!--> <!-->0.002) of calcified lesions along vascular access route. Iliac artery tortuosity differed significantly between groups, with higher angulation observed in patients with technical failure, e.g., bilateral angle sum of iliac bifurcation: 67.6<!--> <!-->±<!--> <!-->21.2 vs. 87.8<!--> <!-->±<!--> <!-->24.3 (p<0.001). PA anatomical variants, dominance patterns, and CTA-/CBCT-based density measurements showed no significant correlation with technical outcome (all p>0.05).</div></div><div><h3>Conclusion</h3><div>Pre- and periprocedural imaging offers valuable predictors of technical success in PAE. Atherosclerotic plaque burden and iliac artery tortuosity serve as practical markers for technical failure, thereby aiding patient selection and procedural planning.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 1","pages":"Pages 121-133"},"PeriodicalIF":3.9,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145314107","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-01DOI: 10.1016/j.acra.2025.10.004
Huairong Zhang , Lina Miao , Li Ma , Xiao Sun , Li-Na Ouyang , Yang Jing , Yifan Wang , Xiang Wang , Pei Wang , Li Zhu
Rationale and Objectives
Accurate prediction of the invasiveness of early-stage pulmonary adenocarcinoma presenting as ground-glass nodules (GGNs) remains highly challenging. This study aims to integrate radiomics features from non-contrast CT (NECT) and contrast-enhanced CT (CECT), deep learning features, and intratumoral habitat features to improve prediction accuracy and provide robust support for clinical personalized surgical decision-making.
Materials and Methods
This dual-center retrospective study included 516 patients with pathologically confirmed GGNs (≤30 mm) from December 2018 to September 2023. Patients from center 1 were randomly divided into training (276 patients) and internal-validation (120 patients) sets, while patients from center 2 were all included into external-validation (120 patients) set. Intratumoral habitat analysis (ITH) was performed on NECT and CECT images using the K-means clustering algorithm. Radiomic features were extracted from the lesion regions and clustered subregions, deep learning features were obtained via a fine-tuned ResNet50 model. After feature selection, eight predictive models were established. Additionally, a dynamic nomogram (the comprehensive model) was developed and subjected to explainable analysis using SHAP (SHapley Additive exPlanations). Model performance was assessed using area under the curve (AUC), decision curve analysis (DCA), and calibration curves.
Results
Among eight predictive models, the comprehensive model, which utilized multi-modal data as input demonstrated the highest accuracy in distinguishing invasive adenocarcinoma (IAC) from pre-invasive lesions (AAH/AIS/MIA). In the training set, the AUC was 0.92 (95% CI: 0.89–0.95), with 84% accuracy, 85% sensitivity, and 84% specificity. In the internal-validation set, the AUC was 0.90 (95% CI: 0.86–0.95), with 82% accuracy, 88% sensitivity, and 74% specificity. In the external-validation set, the AUC was 0.85 (95% CI: 0.80–0.91), with 80% accuracy, 80% sensitivity, and 80% specificity. DCA analysis showed that the nomogram provided the highest net benefit when the threshold probability was ≥0.4, and the Hosmer-Lemeshow test confirmed good calibration (P>0.05). SHAP analysis and the selected of optimal features revealed that wavelet-based texture features, deep learning features, and ITH features made significant contributions to the model's performance.
Conclusion
The comprehensive model (radiomics, deep learning, ITH, clinical variables) enables reliable prediction of the invasiveness of GGNs-ADC. It bridges imaging and pathology, potentially advancing personalized surgical decision-making in early-stage lung adenocarcinoma.
{"title":"Explainable Machine Learning for Predicting Invasiveness of Pulmonary Adenocarcinoma Presenting as Ground-Glass Nodules Using CT Images","authors":"Huairong Zhang , Lina Miao , Li Ma , Xiao Sun , Li-Na Ouyang , Yang Jing , Yifan Wang , Xiang Wang , Pei Wang , Li Zhu","doi":"10.1016/j.acra.2025.10.004","DOIUrl":"10.1016/j.acra.2025.10.004","url":null,"abstract":"<div><h3>Rationale and Objectives</h3><div>Accurate prediction of the invasiveness of early-stage pulmonary adenocarcinoma presenting as ground-glass nodules (GGNs) remains highly challenging. This study aims to integrate radiomics features from non-contrast CT (NECT) and contrast-enhanced CT (CECT), deep learning features, and intratumoral habitat features to improve prediction accuracy and provide robust support for clinical personalized surgical decision-making.</div></div><div><h3>Materials and Methods</h3><div>This dual-center retrospective study included 516 patients with pathologically confirmed GGNs (≤30<!--> <!-->mm) from December 2018 to September 2023. Patients from center 1 were randomly divided into training (276 patients) and internal-validation (120 patients) sets, while patients from center 2 were all included into external-validation (120 patients) set. Intratumoral habitat analysis (ITH) was performed on NECT and CECT images using the K-means clustering algorithm. Radiomic features were extracted from the lesion regions and clustered subregions, deep learning features were obtained via a fine-tuned ResNet50 model. After feature selection, eight predictive models were established. Additionally, a dynamic nomogram (the comprehensive model) was developed and subjected to explainable analysis using SHAP (SHapley Additive exPlanations). Model performance was assessed using area under the curve (AUC), decision curve analysis (DCA), and calibration curves.</div></div><div><h3>Results</h3><div>Among eight predictive models, the comprehensive model, which utilized multi-modal data as input demonstrated the highest accuracy in distinguishing invasive adenocarcinoma (IAC) from pre-invasive lesions (AAH/AIS/MIA). In the training set, the AUC was 0.92 (95% CI: 0.89–0.95), with 84% accuracy, 85% sensitivity, and 84% specificity. In the internal-validation set, the AUC was 0.90 (95% CI: 0.86–0.95), with 82% accuracy, 88% sensitivity, and 74% specificity. In the external-validation set, the AUC was 0.85 (95% CI: 0.80–0.91), with 80% accuracy, 80% sensitivity, and 80% specificity. DCA analysis showed that the nomogram provided the highest net benefit when the threshold probability was ≥0.4, and the Hosmer-Lemeshow test confirmed good calibration (P>0.05). SHAP analysis and the selected of optimal features revealed that wavelet-based texture features, deep learning features, and ITH features made significant contributions to the model's performance.</div></div><div><h3>Conclusion</h3><div>The comprehensive model (radiomics, deep learning, ITH, clinical variables) enables reliable prediction of the invasiveness of GGNs-ADC. It bridges imaging and pathology, potentially advancing personalized surgical decision-making in early-stage lung adenocarcinoma.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 1","pages":"Pages 266-280"},"PeriodicalIF":3.9,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145356767","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号