Interpretable deep learning radiomics from

objectiveTo develop and validate interpretable models integrating standardized uptake value (SUV), radiomics (Rad), and deep learning (DL) features from 18F-FDG PET/CT for differentiating diffuse large B-cell lymphoma (DLBCL) and follicular lymphoma (FL).

methodsThis retrospective study included 250 patients from two centers. Volumes of interest (VOIs) were delineated on PET images for SUV, Rad, and DL features extraction. Feature selection was performed using the Mann–Whitney U test, random forest–based recursive feature elimination, and the least absolute shrinkage and selection operator (LASSO). Seven machine learning classifiers were applied to construct diagnostic models, and fused Rad and DL features were further integrated to construct deep learning radiomics (DLR) models. Model interpretability was assessed using SHapley Additive exPlanations (SHAP). Model performance was evaluated in terms of discrimination, calibration, and clinical applicability.

resultsThe DLR model achieved the best diagnostic performance, with an area under the curve (AUC) of 0.905 and an accuracy of 0.813 in the testing cohort. SHAP analysis identified the Rad feature “original_Maximum” as the most influential predictor for differentiating DLBCL from FL. Calibration curve and decision curve analyses further supported the superiority of the DLR model.

conclusionRad and DL features derived from 18F-FDG PET/CT enable effective differentiation between DLBCL and FL. The proposed SHAP-based interpretable model offers superior diagnostic accuracy and potential clinical utility.