Composite-metric deep learning for pathological complete response prediction from breast MRI

• Conduct a systematic comparison of deep learning models (VGG and ResNet) trained on axial and sagittal 2D breast DCE-MRI slices acquired at T1 and T2 stages from the I-SPY1 dataset. • Propose the GPS as a composite evaluation metric to assess model performance in a clinically meaningful way, integrating multiple metrics into a single composite measure. • Explore the effect of GPS weighting strategies and validate their stability through statistical analysis, enabling robust recommendations on model selection, imaging protocol, and time point configuration. Neoadjuvant chemotherapy (NACT) is commonly administered before surgery to reduce breast tumor size. Achieving a pathological complete response (pCR) after NACT is associated with improved prognosis. Early prediction of pCR help minimize unnecessary treatment related toxicity and facilitates personalized therapeutic strategies. This study introduces a deep learning approach for pCR prediction using 2D breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) data from the public I-SPY1 dataset. We focused on the pretreatment (T1) and early treatment (T2) time points and analyzed both the axial and sagittal planes. Variants of the VGG and ResNet architectures were trained and assessed. To provide a more comprehensive evaluation of the model’s performance beyond accuracy alone, we developed a novel composite metric called the Global Performance Score (GPS). This metric combines accuracy, specificity, and sensitivity through a weighted summation. One-way ANOVA ( F = 1.83, p = 0.127) confirmed GPS robustness across weighting schemes, with (0.4, 0.3, 0.3) chosen as the optimal configuration. Our findings show that VGG-11 and VGG-13 achieved the highest mean GPS scores (92.17 and 92.10, respectively), outperforming ResNet variants. In addition, sagittal views showed a significantly higher average GPS (95.08) compared to axial views (87.12). Similarly, the T2 time point outperformed the T1 time point (GPS: 91.32 vs. 90.89). These results highlight the effectiveness of deep learning in predicting pCR and support the GPS as a robust and interpretable evaluation metric for multi-objective assessment in breast cancer imaging.