A mammographic habitat–clinical fusion model for the preoperative prediction of ductal carcinoma in situ with microinvasion

Ductal carcinoma in situ (DCIS) represents a preinvasive stage of breast cancer, whereas DCIS with microinvasion (DCIS-Mi) has greater metastatic potential and necessitates distinct therapeutic strategies. The precise preoperative differentiation of these conditions remains a critical challenge in clinical practice. This study aimed to develop and validate a novel multimodal model integrating mammographic habitat heterogeneity analysis with clinical parameters for distinguishing DCIS from DCIS-Mi preoperatively. In this multicentre retrospective analysis, 255 patients with pathologically confirmed breast lesions (248 unilateral and 7 bilateral) were included. The cohort was stratified into training (70%) and internal validation (30%) sets, with two external test cohorts used for generalizability assessment. Mammographic ROIs on craniocaudal (CC) and mediolateral oblique (MLO) views were annotated by two board-certified radiologists. A habitat-driven radiomics pipeline yielding 833 quantitative features characterizing tumour heterogeneity and peritumoral microenvironment attributes was applied. LASSO regression with 10-fold cross-validation was used to select the optimal features. Three machine learning algorithms (logistic regression, LightGBM, and ExtraTrees) were compared. All lesions from the same patient were assigned to the same cohort to prevent information leakage. The ExtraTrees classifier demonstrated superior discriminative performance on the internal validation set (AUC: 0.822; 95% CI: 0.694–0.950), significantly outperforming the logistic regression (AUC: 0.649; p < 0.01) and LightGBM (AUC: 0.705; p = 0.03) models. Integrating clinical variables with habitat features in a fusion model enhanced performance, yielding AUCs of 0.831 (95% CI: 0.704–0.957) on the internal validation set and 0.907 (95% CI: 0.816–0.997) on the external Test 2 cohort. In this study, we developed a framework that combines mammographic habitat signatures with clinical metadata for improved preoperative risk stratification of DCIS-Mi. The model demonstrates promising performance and interpretability; however, prospective validation is required before clinical translation.