An interpretable contrast-enhanced CT radiomics-based pipeline incorporating automatic segmentation for predicting...

Non-invasive assessment of clear cell renal cell carcinoma (ccRCC) International Society of Urological Pathology grade group (ISUP GG) remains a clinical challenge. This study aims to develop an interpretable radiomics-based pipeline incorporating automatic segmentation for predicting the ISUP GG in ccRCC.

This retrospective study included 276 pathologically confirmed ccRCC patients who underwent preoperative contrast-enhanced CT (193 train, 83 test), and were stratified by ISUP GG (1-2 vs. 3-4). Radiomics features were extracted from the auto‑segmented tumor regions on corticomedullary-phase images. Feature selection was performed using Recursive Feature Elimination Cross-Validation across 5 classifiers. Along with clinical information, we ultimately constructed radiomics, clinical, and combined models. Models comprehensively evaluated via multiple metrics, with partial results visualized through Receiver Operating Characteristic and calibration curves, and decision curve analysis, and further interpretability analysis conducted.

The combined model significantly outperformed the clinical and radiomics models, achieving an AUC of 0.907 (95% CI 0.868-0.947) vs. 0.802 (95% CI 0.735-0.869) and 0.769 (95% CI 0.702-0.836), both p < 0.001. In the test, the combined model significantly outperformed the clinical model, achieving an AUC of 0.834 (95% CI 0.737-0.931) vs. 0.698 (95% CI 0.575-0.821), p = 0.001, but showed no significant difference compared with the radiomics model (AUC 0.834 [95% CI 0.737-0.931] vs. 0.796 [95% CI 0.695-0.897], p = 0.328). Calibration curves and DCA indicated that the combined model demonstrated good calibration and better clinical net benefit.

The interpretable pipeline may have potential as a non-invasive tool for predicting ISUP GG in ccRCC.