Optimizing Deep Learning for Renal Mass Characterization in Challenging Cases: A Comparative Study of Spatial-Input Strategies and a Validated ROI-Only DLR Nomogram.

PURPOSE: Our purpose was to develop and validate an integrated clinical deep learning radiomics (DLR) nomogram for differentiation of fat-poor angiomyolipoma (fp-AML) from clear-cell renal cell carcinoma (ccRCC) in diagnostically challenging cases and to systematically compare spatial-input strategies for optimal model performance.

METHODS

This retrospective study enrolled 469 patients with pathologically confirmed renal masses (fp-AML or ccRCC). A stratified cohort partitioning strategy was employed: the training cohort comprised cases with consistent radiological-pathological diagnoses, and the independent test cohort was purposely enriched with "gray-zone" cases (indeterminate imaging or prior discordant interpretations). Four spatial-input architectures for DLR feature extraction were systematically compared: region of interest (ROI) only, 1mm ROI expansion, uncropped full-slice, and conventional cropping. After rigorous feature selection, seven machine learning algorithms were evaluated to identify the optimal model. A nomogram was constructed and validated using receiver operating characteristics analysis, calibration curves, and decision curve analysis.

RESULTS

The ROI-only DLR model demonstrated superior discriminative performance among all spatial-input strategies. In the independent "stress-test" cohort, it achieved an area under the curve (AUC) of 0.820, significantly outperforming standalone radiomics (AUC 0.735), deep learning (AUC 0.778), and other DLR variants (AUCs 0.736-0.785). The nomogram further improved diagnostic performance with an AUC of 0.840. Decision curve analysis confirmed the superior net benefit of the nomogram across a wide range of threshold probabilities, and calibration curves demonstrated good agreement between predicted and observed outcomes.

CONCLUSION

The ROI-only DLR nomogram, specifically validated on diagnostically challenging cases, provides a robust and non-invasive decision-support tool for differentiating fp-AML from ccRCC, potentially reducing unnecessary surgeries and improving preoperative management of indeterminate renal masses.