Predicting Renal Sinus Invasion in Renal Cell Carcinoma Using a Deep Learning-Assisted Radiomics Model: A Comparison with Human Assessment

Background This study aimed to develop a deep learning-assisted radiomics model (PRSI-model) to improve preoperative prediction of renal sinus invasion (RSI) in renal cell carcinoma (RCC). Methods A retrospective analysis was conducted on 437 patients (91 with RSI) who underwent nephrectomy across two centers from 2013 to 2018. Utilizing a ResNeXt-based convolutional neural network, the model extracted imaging features and integrated clinical parameters to optimize predictive performance. Results Trained on a primary cohort (N = 276), the model underwent internal (N = 70) and external (N = 91) validation. The PRSI-model alone achieved an AUC of 0.82, which improved to 0.90 when combined with clinical factors such as tumor exogeneity and distance from the renal collecting system (D_RPCS). Subgroup analyses revealed superior performance in non-clear cell RCC (AUC 0.81) compared to clear cell subtypes (AUC 0.65). When evaluated against assessments by 10 clinical experts (radiologists and urologists), the model outperformed human accuracy (model AUC: 0.82 vs. average human AUC: 0.59), particularly in high-complexity cases (higher R.E.N.A.L scores) and heterogeneous tumors. Conclusion These findings highlight the PRSI-model’s potential as a non-invasive tool for preoperative RSI prediction, aiding surgical decision-making. Future studies should validate the model across diverse populations and refine its efficacy for specific histologic subtypes to enhance clinical applicability.