Interpretable CT habitat analysis for preoperative prediction of histological differentiation in gastric cancer.

PURPOSE: To develop and validate an interpretable preoperative model for predicting histological differentiation in gastric cancer using CT habitat analysis combined with clinical features. METHODS: This retrospective study included 978 patients with pathologically confirmed gastric cancer from two institutions. Patients from Center 1 (n = 678) were randomly divided into a training cohort (n = 406) and a testing cohort (n = 272), while patients from Center 2 (n = 300) served as an external validation cohort. Independent clinical predictors of histological differentiation were identified using multivariable logistic regression to construct a clinical model. Tumor habitats were generated from preoperative venous-phase CT using K-means clustering, and habitat scores, subregional volumes, and volume fractions were extracted to develop a Habitat model. Clinical and habitat features were then integrated into a nomogram. RESULTS: Age, sex, tumor location, CT-detected T stage, CT-detected N stage, and CA19-9 level were independent predictors of histological differentiation. The optimal habitat partition was achieved with two clusters (k = 2) according to the Calinski-Harabasz index. In the training cohort, area under the receiver operating characteristic curve (AUC) were 0.747 (95% CI: 0.700-0.794) for the Clinical model, 0.835 (0.794-0.874) for the Habitat model, and 0.872 (0.836-0.904) for the Nomogram model. In the testing cohort, corresponding AUCs were 0.652 (0.585-0.716), 0.778 (0.723-0.831), and 0.803 (0.745-0.855). In the external validation cohort, AUCs were 0.682 (0.620-0.738), 0.831 (0.780-0.881), and 0.839 (0.791-0.887), respectively. SHapley Additive exPlanations (SHAP) analysis identified habitat-derived features as the strongest contributors to model predictions. CONCLUSION: CT habitat analysis enables accurate preoperative prediction of histological differentiation in gastric cancer. Incorporating clinical variables further improves performance.