MRI-based radiomics with SHAP interpretation for preoperative prediction of upstaging in ductal carcinoma in situ: a comparative study of intratumoral, peritumoral, habitat, and fusion models

Background: Preoperative prediction of upstaging in ductal carcinoma in situ (DCIS) is crucial to avoid unnecessary sentinel lymph node biopsy (SLNB) in low-risk patients. We aimed to develop and validate an interpretable magnetic resonance imaging (MRI)-based radiomics model for this purpose and to systematically compare the predictive value of intratumoral, peritumoral, and habitat-based features. Methods: This retrospective study included 108 women with biopsy-proven DCIS. Radiomics features were extracted from intratumoral and peritumoral regions (2, 4, 6 mm) on dynamic contrast-enhanced MRI. Six models (clinical, intratumoral, peritumoral, habitat, feature-fusion, image-fusion) were developed and compared using multiple machine learning classifiers. The best-performing model was validated on an independent test set (n=33). Model interpretability was achieved using SHapley Additive exPlanations (SHAP). Results: The image-fusion model integrating intratumoral and 2-mm peritumoral features showed promising performance, with an area under the curve (AUC) of 0.892 [95% confidence interval (CI): 0.817-0.967] in the training set and 0.864 (95% CI: 0.735-0.992) in the test set. SHAP analysis identified textural heterogeneity as a key predictor. Using predefined radiomics score (Rad-score) thresholds derived from the training set, the high-sensitivity threshold achieved a negative predictive value of 100% (10/10) and the high-specificity threshold achieved a positive predictive value of 61.5% (8/13) in the independent test set. An exploratory ultra-low-risk threshold (Rad-score <0.25) identified a subgroup of 5 out of 33 patients (15.2%) with no upstaging. Conclusions: The proposed MRI-based radiomics model may assist in preoperative risk stratification for DCIS upstaging, but these findings are exploratory and require prospective multicenter validation.