AI-based diagnosis of clear-cell renal cell carcinoma based on non-contrast CT

The accurate characterization of renal tumors, particularly clear-cell renal cell carcinoma (ccRCC), traditionally requires contrast-enhanced computed tomography (CT) (CECT), which is contraindicated in many patients. Noncontrast CT (NCCT) is widely accessible but is considered limited for tumor analysis. This study aims to develop and validate a deep learning system for automated renal tumor localization and ccRCC classification using only NCCT. In a multicenter retrospective study of 1,902 patients with paired NCCT and CECT series, expert annotations were propagated to NCCT via 3D hybrid affine and deformable registration. An integrated AI pipeline combining a 3D U-Net for segmentation with three 3D classifiers (ResNet-50, Vision Transformer, Swin Transformer) was developed. Three model configurations (NCCT-only, CECT-only, multiseries fusion) were evaluated using area under the curve (AUC), accuracy, sensitivity, specificity, and DeLong’s test. The NCCT-based segmentation model achieved tumor Dice coefficients approaching the CECT-based model (0.89 ± 0.08 vs. 0.96 ± 0.05 internally). For classification, NCCT models showed performance that was noninferior to that of CECT models, with no significant AUC differences (internal: 0.80 vs. 0.78 for ResNet-50). On the external test set ( n = 231), NCCT AUCs were 0.80–0.82, which was comparable to CECT models. The fusion model outperformed the NCCT model using ResNet-50, which consistently surpassed transformer-based architectures. This study demonstrates an AI system for accurate renal tumor localization and ccRCC classification directly from NCCT, achieving performance comparable to CECT-based models. This contrast-free approach could enhance accessible renal cancer screening and provide diagnostic support for patients with contraindications to contrast media or in resource-limited settings.