A Three-Slice Deep Learning-Radiomics Nomogram for Challenging Renal Mass Cases: A Double-Center Study.

To determine the optimal spatial-depth architecture for deep learning-radiomics (DLR) differentiation of fat-poor angiomyolipoma (fp-AML) from clear cell renal cell carcinoma (ccRCC), and to validate the resulting integrated nomogram under stringent stress test conditions using only diagnostically equivocal or discordant cases, thereby simulating the real-world clinical scenarios where such a tool is most needed, this retrospective double‑center study enrolled 583 patients (329 training, 131 internal stress test, and 123 external stress test). Two DLR pipelines were compared: a 2D single-slice input (the maximal tumor cross-section) versus a 2.5D three-slice input that stacked the maximal axial slice together with its immediate superior and inferior adjacent slices. Both pipelines used a ResNet-50 backbone with fused radiomic features. The optimal 2.5D DLR signature was integrated with independent clinical predictors into a nomogram. Validation cohorts comprised exclusively cases with indeterminate imaging or imaging-pathology discordance. The 2.5D DLR model outperformed the 2D DLR model in both the internal stress test cohort (AUC 0.821 vs. 0.782) and the external validation cohort (AUC 0.809 vs. 0.763). The integrated nomogram achieved the highest discriminative performance (internal AUC 0.855; external AUC 0.823). Decision curve analysis confirmed superior net clinical benefit across both stress test cohorts, and calibration curves demonstrated good agreement between predicted and observed outcomes. The 2.5D architecture represents the optimal deep learning strategy for differentiating fp-AML from ccRCC. The integrated nomogram provides a potential decision support tool purpose-built for resolving diagnostic uncertainty in the most challenging real-world scenarios, with robust performance validated across two independent stress test cohorts.