An artificial intelligence system for early prediction of breast cancer regression pattern during neoadjuvant chemotherapy.

3 Background: Breast cancer exhibits diverse tumor regression patterns (TRP) following neoadjuvant chemotherapy (NAC). We aims to develop an artificial intelligence (AI) system utilizing longitudinal magnetic resonance imaging (MRI) for precise TRP prediction. Methods: This retrospective study involved 2249 breast cancer patients from 12 institutions. The dataset comprised a training cohort (n=1006) from institutions I and II, and an external validation cohort (n=1243) from institutions III-XII. We utilized a 3D U-Net model for automated tumor delineation, incorporating spatial habitat analysis and 3D ResNet-50 model for extracting imaging featur, to develop an AI system. Results: The 3D U-Net model showed significant accuracy, with Dice coefficients of 0.875 in the validation cohort. The AI system achieved areas under the curve of 0.912. It demonstrated robust performance across diverse molecular subtypes (accuracies: 80·68% to 87·87%) and tumor stages (accuracies: 80·16% to 84·98%) in the validation cohorts. Conclusions: Our study provides a noninvasive AI system for early TRP prediction in breast cancer, potentially assisting clinicians in adjusting NAC regimens and planning breast-conserving surgery.