Predicting Pathologic Complete Response in Breast Cancer Using Pre-Treatment DCE-MRI

Prediction of pathological complete response (pCR) in breast cancer is relevant for clinical decision making and individual risk assessment, but remains challenging at present. This study aimed to design and validate clinically-informed, interpretable dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) framework that combines quantitative imaging biomarkers with clinical predictors. The retrospective computational analysis involved 129 development and 24 held-out test cases. Pretreatment DCE-MRI parameters, such as tumor volume, early enhancement, late enhancement and washout characteristics, were derived. To compare the clinical-only logistic regression baseline (B01) with the regularized clinical-DCE fusion model (B02), repeated stratified cross-validation, locked-model held-out evaluation, paired bootstrapping, feature-group ablation and coefficient-based interpretation were used. During internal validation, B02 improved ROC-AUC from 0.673 to 0.726 and PR-AUC from 0.415 to 0.524. These improvements were not replicated in held-out testing, however, as B02 had lower ROC-AUC (0.655) than in training (0.672) and PR-AUC (0.407) than in training (0.461). Conversely, B02 improved balanced accuracy (0.622 versus 0.563), specificity (0.529 versus 0.412), and F1-score (0.500 versus 0.455), while reducing the Brier score (0.230 versus 0.268). The imaging feature contribution was heterogeneous as determined by ablation analysis, suggesting potential redundancy in imaging features and potentially limited generalization. These observations suggest that quantitative DCE-MRI can give additional information on threshold dependent classification and probabilistic prediction but not necessarily better discrimination. Larger scale independent multicentre studies and prospective decision-utility assessments are needed before clinical implementation.