治疗中期Delta MRI影像组学提高乳腺癌病理完全缓解的敏感度与风险分层：多中心研究

RATIONALE AND OBJECTIVES: Predicting pathological complete response (pCR) to neoadjuvant chemotherapy in breast cancer remains challenging. We developed and validated mid-treatment MRI delta radiomics models for pCR prediction.

MATERIALS AND METHODS: This multicenter retrospective study included 242 patients (training: n = 126; internal validation: n = 55; external validation: n = 61). Delta radiomics features (mid-treatment minus pre-treatment) were extracted from T1-weighted contrast-enhanced sequences following intraclass correlation coefficients filtering (ICC>0.75). Six models combining clinical features with radiomics were developed using Random Forest with nested cross-validation (5×10 folds). Model performance was assessed using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and risk stratification.

RESULTS: The combined clinical-delta radiomics model achieved external validation AUC of 0.747 (95% CI: 0.615-0.863), comparable to clinical-only model (AUC=0.777, P = 0.557). Radiomics-only models showed limited generalization (AUC=0.627-0.698). The combined clinical-delta model demonstrated significantly higher sensitivity (90.6% vs 65.6%, P = 0.022) and effective risk stratification: predicted responder group (>0.7 predicted pCR probability) achieved 76.2% pCR rate versus 25.0% in predicted non-responder group (<0.3 predicted pCR probability), representing 51%-point separation (P = 0.003). Cancer subtype was the dominant predictor (44.3% relative importance), with nine delta radiomics features collectively contributing 55.7%.

CONCLUSIONS: Mid-treatment delta MRI radiomics combined with clinical features achieves high sensitivity (90.6%) and effective risk stratification for pCR prediction. While AUC improvement over clinical features alone was not statistically significant, the enhanced sensitivity and risk stratification support potential clinical utility, warranting larger prospective validation.