Predicting aggressiveness of clear cell renal cell carcinoma via mri using artificial intelligence: implications for surgical planning in a retrospective multicenter study

Abstract Surgical planning for clear cell renal cell carcinoma (ccRCC) is highly dependent on tumor aggressiveness. However, current decisions are predominantly based on postoperative pathology, limiting opportunities for timely and personalized treatment. A reliable, non-invasive, preoperative method for assessing tumor grade could enhance surgical outcomes and avoid unnecessary interventions. This study aimed to develop and validate an artificial intelligence (AI) model based on preoperative magnetic resonance imaging (MRI) to predict ccRCC aggressiveness and guide surgical decision-making. We retrospectively analyzed 288 preoperative MRI volumes from 332 ccRCC patients across three medical centers in China (May 2018 - March 2025). All MRI scans included axial T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) maps. Of these, 153 cases lacked lesion location and pathology annotations and were used to pre-train a foundation model. The remaining 96 cases, which included lesion location and pathology results information, were utilized for training and internal validation through cross-validation. To evaluate the model’s generalizability, an independent external validation set comprising 39 cases from two additional centers was used. A vision transformer (ViT) based foundation model and classification model were trained to automatically predict the aggressiveness of ccRCC based on preoperative MRI images. In internal validation, the AI model achieved an ROC AUC of 0.786 (95% CI: 0.682–0.877), PR AUC of 0.749 (95% CI: 0.602–0.861), with 71.4% sensitivity and 70.5% specificity in predicting high-grade ccRCC. In external validation, it reached an ROC AUC of 0.935, demonstrating improved performance compared with R.E.N.A.L. nephrometry score-based nomograms (AUC: 0.691 internal, 0.796 external). The model also yielded higher PR AUCs than R.E.N.A.L. scores (0.532 internal, 0.527 external), and provided interpretable outputs aligned with surgical decision-making, supporting choices between nephron-sparing surgery and radical nephrectomy. This study demonstrates that an AI-based, MRI-driven framework can accurately predict tumor aggressiveness in ccRCC before surgery. The model holds promise for integrating into preoperative workflows, supporting precision surgery, and improving patient outcomes in renal oncology.