Biphasic CT clustering-based habitat radiomics predicts WHO/ISUP nuclear grade in clear cell renal cell carcinoma.

PURPOSE: To evaluate a biphasic CT clustering-based habitat radiomics approach for predicting the World Health Organization/International Society of Urological Pathology (WHO/ISUP) nuclear grade in clear cell renal cell carcinoma (ccRCC) and assess the discriminatory value of different tumor subregions. METHODS: This retrospective study collected 473 ccRCC patients (training: n = 337; test: n = 136). Thin-section corticomedullary phase (CMP) and excretory phase (EP) CT images were preprocessed and registered. NnU-Net was used for automated segmentation of tumor. Tumor habitats were generated via k-means clustering of three parameters: CMP attenuation values (CMP_HU), EP attenuation values (EP_HU), and their interphase difference (ΔHU). Radiomics features were extracted from whole-tumor volumes, individual subregions, and subregional combinations, followed by feature selection and model development. RESULTS: Five distinct subregions were identified. Subregion 4 ( representing the tumor viable-necrosis transition zone ) -based Random Forest model achieved diagnostic performance statistically comparable to whole-tumor-based model in both CMP ( AUC: 0.773 vs. matched whole-tumor AUC 0.819, DeLong' p = 0.152 ) and EP ( AUC: 0.810 vs. matched whole-tumor AUC 0.827, DeLong' p = 0.593 ). Subregion 2 ( necrotic core ) -based model showed similar efficacy to whole-tumor model in EP ( AUC: 0.771 vs. matched whole-tumor AUC 0.802, DeLong' p = 0.219 ). Multi-subregion models containing Subregion 4 maintained statistically comparable performance with whole-tumor models, with Subregion 1+4 demonstrating optimal performance. CONCLUSION: Biphasic CT clustering-based habitat analysis reveals that specific tumor subregions-particularly the tumor viable-necrosis transitional zone-can approximate whole-tumor diagnostic performance for ccRCC nuclear grading. This method provides spatial mapping of tumor heterogeneity through visualizable CT subregions, offering a potential complementary tool for non-invasive ccRCC characterization. Further prospective multicenter validation with histopathological correlation is needed before clinical implementation.