Optimizing Deep Learning for Renal Mass Characterization in Challenging Cases: A Comparative Study of Spatial-Input...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Differentiating fat-poor angiomyolipoma (fp-AML) from clear-cell renal cell carcinoma (ccRCC) is challenging due to overlapping imaging features.
• Accurate preoperative diagnosis is crucial for appropriate patient management and avoiding unnecessary surgeries.

Purpose of the Study:

• To develop and validate a deep learning radiomics (DLR) nomogram for distinguishing fp-AML from ccRCC in difficult cases.
• To compare different spatial-input strategies for optimizing DLR model performance.

Main Methods:

• Retrospective analysis of 469 patients with pathologically confirmed renal masses (fp-AML or ccRCC).
• Development of DLR models using four spatial-input strategies: ROI-only, 1mm ROI expansion, uncropped full-slice, and conventional cropping.
• Validation using receiver operating characteristics analysis, calibration curves, and decision curve analysis on a test cohort enriched with challenging cases.

Main Results:

• The ROI-only DLR model showed superior performance, achieving an AUC of 0.820 in the independent test cohort.
• The integrated nomogram further improved diagnostic performance with an AUC of 0.840.
• The nomogram demonstrated significant net benefit and good calibration, outperforming standalone radiomics and deep learning models.

Conclusions:

• The developed ROI-only DLR nomogram is a robust, non-invasive decision-support tool for differentiating fp-AML from ccRCC.
• This tool can improve preoperative management of indeterminate renal masses and potentially reduce unnecessary surgeries.