Habitat‐based radiomic model for predicting muscle invasion in bladder cancer: A multi‐center study using enhanced‐CT and machine learning

BACKGROUND: Accurate assessment of muscle invasion in bladder cancer is crucial for guiding treatment and prognosis. Habitat-based radiomics, which accounts for tumor heterogeneity, may enhance evaluation of tumor status and outcomes. PURPOSE: This research primarily investigates the efficacy of a novel habitat-based radiomic model in predicting muscle invasion in bladder cancer. METHODS: We retrospectively analyzed 325 bladder cancer patients from two institutions (July 2018-July 2023). Patients were divided into a training cohort (231 cases, Institution 1) and an external test cohort (94 cases, Institution 2). CT images were standardized, and areas of interest (AOIs) were delineated. Nineteen texture features were extracted from each AOI, and K-means clustering identified intratumoral habitats. Radiomic features from each habitat were extracted using PyRadiomics and used to build a habitat model with the ExtraTree algorithm. For comparison, we also developed uniphase, multiphase, and clinical models. Model performance was evaluated by sensitivity, specificity, accuracy, and area under the ROC curve (AUC). The Delong test compared diagnostic performance between models. RESULTS: Three distinct habitats were identified within bladder tumors. The habitat model achieved an AUC of 0.947 (95% CI: 0.911-0.982) in the training cohort and 0.825 (95% CI: 0.704-0.946) in the external test cohort. In the training cohort, the habitat model outperformed the uniphase (p = 0.003), multiphase (p = 0.036), and clinical models (p = 0.049). The combined habitat and clinical model showed superior diagnostic performance compared to uniphase (p = 0.019) and multiphase clinical (p = 0.069) fusion models. The radiomics signature integrating habitat and multiphase features reliably predicted muscle invasion across the entire cohort (AUC = 0.922, 95% CI: 0.883-0.960). CONCLUSIONS: Habitat-based radiomic features combined with machine learning enable accurate preoperative prediction of muscle invasion in bladder cancer using CT images.