An Interpretable Deep Learning Framework for Discovery and Clinical Validation of Deep Radiomic Signatures in Tumor Classification

The paper presents an interpretable imaging-signature discovery framework that combines tumor segmentation, Grad-CAM-guided region localization, mutual-information-based patient-specific thresholding, downstream deep classification, and SHAP analysis of radiomic features. It evaluates the approach on the public BUSI breast ultrasound, KiTS renal CT, and BraTS brain tumor datasets, plus a private UF Health renal CT cohort. The authors report that signature-region radiomics outperform conventional whole-tumor radiomics in discrimination while offering greater biological interpretability. The abstract does not provide exact metrics, cohort sizes, statistical tests, or external validation details.