Renal Cell Carcinoma Texture Analysis With the KiTS23 Dataset: A Retrospective Study.

Background and Aims: Computed tomography texture analysis, powered by machine learning techniques, may differentiate clear cell renal cell carcinoma (ccRCC) from other renal tumor subtypes, such as papillary and chromophobe variants, or benign renal masses such as oncocytomas, as demonstrated here using the KiTS23 dataset. Methods: After excluding multifocal cases to avoid lesion-level labeling ambiguity and intrapatient lesion heterogeneity, 396 cases were included. Using PyRadiomics, 386 radiomics features were initially extracted from preprocessed computed tomography volumes and tumor segmentation masks. Subsequent multistage feature selection reduced the candidate feature set to 73 radiomics descriptors, and LASSO further selected six features for the primary classifier. Candidate machine-learning models were evaluated, with class-weighted Logistic Regression selected as the primary model. Threshold adjustment using the Youden index and bootstrap feature-selection stability analysis were performed. Results: The LASSO-selected class-weighted Logistic Regression model showed the best overall performance. Using the default probability threshold of 0.5, it achieved an accuracy of 76.5%, an area under the receiver operating characteristic curve of 80.4%, a precision of 91.2%, a sensitivity of 73.8%, and a specificity of 82.9% on the held-out test set. Youden-index thresholding increased sensitivity but reduced specificity; therefore, the fixed threshold of 0.5 was retained as the primary operating threshold. Bootstrap stability analysis showed that several LASSO-selected radiomics descriptors were repeatedly selected across resampled training cohorts. Conclusions: Radiomics-based machine learning shows promise for discriminating ccRCC from non-ccRCC renal tumors, but the results remain preliminary and have been evaluated internally. Given the moderate discrimination and lack of external validation, further studies using larger, multicenter datasets with standardized radiomics workflows are necessary.