Magnetic resonance imaging radiomics for predicting pathologic complete response to neoadjuvant chemotherapy in breast cancer.
Authors
Affiliations (1)
Affiliations (1)
- Afyonkarahisar University of Health Sciences, Faculty of Medicine, Department of Radiology - Afyonkarahisar, Turkey.
Abstract
The aim of this study was to develop and validate machine learning models integrating clinical and pre-treatment dynamic contrast-enhanced magnetic resonance imaging radiomic features for predicting pathologic complete response to neoadjuvant chemotherapy in breast cancer. This retrospective study analyzed 299 breast cancer patients from the publicly available Duke Breast Cancer magnetic resonance imaging dataset who received neoadjuvant chemotherapy. From pre-treatment dynamic contrast-enhanced magnetic resonance imaging, 534 IBSI-compliant radiomic features were extracted. A five-method consensus feature selection pipeline retained 24 features. Notably, three prediction models were developed: clinical-only (Model A), radiomic-only (Model B), and combined clinical-radiomic (Model C). In total, seven machine learning algorithms were compared using 20 repeated five-fold stratified cross-validation with SMOTE applied within each training fold and assessed on a held-out test set (30%, n=90). Performance was evaluated using AUC with bootstrap confidence intervals, calibration (Brier score), and decision curve analysis. The pathologic complete response rate was 20.4% (61/299). The combined model (Gradient Boosting) achieved the highest test AUC of 0.723 (95%CI 0.595-0.835), with sensitivity of 0.333 and specificity of 0.833. DeLong tests revealed no statistically significant differences between model pairs (all p>0.05). Decision curve analysis demonstrated the highest net benefit for the combined model across 10-50% threshold probabilities. The combined clinical-radiomic model demonstrated modest discriminative ability for pathologic complete response prediction. The limited sensitivity suggests potential utility as a rule-out test to identify patients unlikely to achieve pathologic complete response. These hypothesis-generating findings require external validation in independent multi-center cohorts.