• Differentiating TFE3-RCC from common RCC subtypes based on CT features remains a formidable challenge. • Interpretable ML models using CT radiomics effectively differentiate TFE3-RCC from common RCC. • This approach can improve preoperative diagnosis for TFE3-RCC, potentially improving patient stratification and treatment outcomes. TFE3-RCC is rare, hard to distinguish from common RCC on CT, posing preoperative diagnostic challenges for clinicians. This two-center study aimed to develop interpretable machine learning models using radiomics to differentiate Xp11.2/TFE3 translocation renal cell carcinoma (TFE3-RCC) from common renal cell carcinoma (RCC) subtypes using computed tomography (CT) images. Methods: Retrospective data from 1394 patients (39 TFE3-RCC, 1355 non-TFE3 RCC) were analyzed. A propensity score matching (PSM) was applied, resulting in 234 cases (TFE3: n=39, non-TFE3: n=195) included in the radiomics study. CT images were segmented using an AI-based model, and 102 radiomic features (shape, first-order statistics, texture) were extracted. Recursive feature elimination (RFE) with random forest and gradient boosting models were used for feature selection and model development. Performance was evaluated via area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. The patients with TFE3-RCC were significantly younger (36.51 ± 12.68 vs. 57.30 ± 12.00 years, P < 0.05), and had more frequent calcification (30.8% vs. 6.4%, P < 0.05) and were larger (5.50 ± 3.17 cm vs. 4.11 ± 2.06 cm, P =0.005) than those with non-TFE3 RCC, and preferred to implicate females (female: 46.2% vs. 29.3%, P = 0.023). The model identified six optimal features, with skewness (relative weight: 44.57%) and first-order statistics as key predictors. The training set and test set achieved stable performances with AUC (0.951 (95% CI: 0.920–0.983) and 0.864 (95% CI: 0.749–0.979)) and accuracy (0.878 and 0.852). Interpretable radiomics-based machine learning models effectively differentiate TFE3-RCC from common RCC subtypes, with skewness and intensity features as critical biomarkers. This approach may improve preoperative diagnosis, though larger multi-center studies and integration of multi-omics data are needed for clinical translation.
Liu et al. (Sun,) studied this question.