Multimodal CT radiomics and 2.5D deep learning for preoperative grading of clear cell renal cell carcinoma in older patients: a multicenter study

Introduction To develop and validate a multimodal CT-based model integrating intra-peritumoral radiomics, 2.5D deep learning (DL), and clinical-radiological features for preoperative prediction of World Health Organization/International Society of Urological Pathology (WHO/ISUP) grade in older patients with clear cell renal cell carcinoma (ccRCC). Methods This multicenter retrospective study included 1,103 patients older than 60 years with pathologically confirmed ccRCC from four hospitals. Patients were allocated to a training cohort (n=459), an internal validation cohort (n=198), and two independent external test cohorts (test1, n=227; test2, n=219). WHO/ISUP grades I–II were classified as low grade and grades III–IV as high grade. Radiomics features were extracted from corticomedullary-phase CT images using intratumoral and progressively expanded intra-peritumoral regions of 1, 2, and 3 mm. In parallel, 2D and 2.5D DL features were extracted using ImageNet-pretrained ResNet34 networks, with the 2.5D approach integrating axial, coronal, and sagittal tumor views. After feature selection, five machine learning classifiers were evaluated. The optimal radiomics model, DL model, and clinical model were integrated using stacking-based decision-level fusion. Model performance was assessed using receiver operating characteristic analysis, calibration curves, decision curve analysis (DCA), and SHapley Additive exPlanations (SHAP). Results The random forest (RF) model based on intra-peritumoral 3-mm radiomics features achieved areas under the receiver operating characteristic curve (AUCs) of 0.754 (95% CI, 0.683–0.825) and 0.725 (95% CI, 0.655–0.794) in the test1 and test2 cohorts, respectively. The RF model based on 2.5D DL features achieved corresponding AUCs of 0.756 (95% CI, 0.682–0.830) and 0.713 (95% CI, 0.643–0.783). The multimodal combined model achieved numerically higher external AUCs than the selected single-modality models, reaching 0.775 (95% CI, 0.705–0.845) in test1 and 0.732 (95% CI, 0.664–0.800) in test2. DCA demonstrated favorable clinical net benefit for the combined model, while SHAP analysis identified the 2.5D DL component as the largest contributor to model predictions. Discussion A multimodal CT-based framework integrating intra-peritumoral radiomics, 2.5D DL, and clinical-radiological information enabled noninvasive preoperative prediction of WHO/ISUP grade in older patients with ccRCC and maintained predictive performance across independent external cohorts. This approach may provide complementary imaging and clinical information for preoperative risk stratification and individualized treatment planning.