An explainable AI system for preoperative prediction of renal sinus fat invasion in clear cell renal cell carcinoma: a multicenter study

This study aimed to develop and evaluate a multimodal predictive system (MPS) for assessment of renal sinus fat invasion (RSFI) in clear cell renal cell carcinoma (ccRCC) patients. A total of 875 ccRCC patients from four independent institutions were retrospectively analyzed. The cohort was divided into a training set ( n = 445), validation set ( n = 111), and two multicentre evaluation sets ( n = 181, n = 138). Radiomics features were extracted from the intratumoral ROI and tumor-plus-peritumoral ROIs expanded by 1, 3, and 5 mm. Multiple machine learning classifiers were constructed and evaluated. In parallel, 2D, 2·5D, and 3D deep learning (DL) models based on ResNet34 were trained. A clinical model was developed using significant variables. The representative models from each modality were integrated via stacking ensemble learning to construct the MPS. Diagnostic performance was evaluated by AUC, sensitivity, specificity, and accuracy, while model interpretability was analyzed using SHAP. The peri-3 mm ExtraTrees radiomics model achieved the highest AUC values of 0.753 (95% CI: 0.6741 - 0.8322) on test1 set and 0.751 (95% CI: 0.6610 - 0.8406) on test2 set. Among DL models, the 2·5D-Peri-5 mm model showed the most consistent numerical performance across the two multicentre evaluation cohorts with AUC values of 0.762 (95% CI: 0.6886 - 0.8360) on test1 set and 0.774 (95% CI: 0.6828 - 0.8651) on test2 set. The clinical model reached AUCs of 0.735 (0.6571 - 0.8121) and 0.736 (0.6499 - 0.8216) in the test sets. The MPS integrating the selected radiomics, DL, and clinical models via an SVM stacking strategy achieved the highest numerical AUCs, reaching 0.791 (95% CI: 0.7213–0.8607) in test1 and 0.777 (95% CI: 0.6894–0.8655) in test2. SHAP analysis revealed that features derived from the 2·5D-Peri-5 mm model contributed most significantly to RSFI prediction. The MPS may provide adjunctive quantitative information for preoperative RSFI risk stratification in patients with CT findings suspicious for RSFI, but prospective evaluation against existing expert assessment is required before clinical use.