Predicting Ki-67 expression levels in non-small cell lung cancer using an explainable CT-based deep learning radiomics model

Objective To predict Ki-67 expression levels in non-small cell lung cancer (NSCLC) using an interpretable model combining clinical-radiological, radiomic, and deep learning features. Methods This retrospective study included 259 NSCLC patients from Center 1 (training/validation sets) and 112 from Center 2 (independent test set). Patients were grouped by a 40% Ki-67 cutoff. Radiomic features and deep learning features were extracted from CT images, where the deep learning features were obtained via a deep residual network (ResNet18). The least absolute shrinkage and selection operator (LASSO) was used to select optimal features and compute radiomics (rad-score) and deep learning (deep-score) scores. Univariate and multivariate logistic regression were used to identify independent clinical-radiological predictors of Ki-67. Four support vector machine models were developed: a clinical-radiological model (based on independent clinical-radiological features), a radiomic model (using the rad-score), a deep learning model (using the deep-score), and a combined model (integrating all the above features). SHapley Additive exPlanations (SHAP) analysis was used to visualize feature contributions. Models’ performance was assessed using receiver operating characteristic (ROC) curves and the integrated discrimination improvement (IDI) index. Results High Ki-67 expression occurred in 76 (42.0%), 38 (48.7%), and 33 (29.5%) patients in the training, validation, and independent test sets, respectively. In the independent test set, the combined model achieved the highest predictive performance, with an AUC of 0.892 (95% CI: 0.828–0.956). This improvement over the clinical-radiological (0.820, 95% CI: 0.721–0.918), radiomics (0.750, 95% CI: 0.655–0.844), and deep learning (0.817, 95% CI: 0.732–0.902) models was statistically significant (all p<0.05), as supported by IDI values of 0.115, 0.288, and 0.095, respectively. SHAP analysis identified the deep-score, histological type, and rad-score as key predictors. Conclusion The interpretable combined model can predict Ki-67 expression in NSCLC patients. This approach may provide imaging evidence to assist clinicians in optimizing personalized therapeutic strategies.