Early prediction of response to neoadjuvant therapy (NAT) is crucial for optimizing breast cancer treatment strategies. Here, we present a novel breast self-supervised temporal learning framework (BSTNet) for predicting the pathological complete response (pCR) using longitudinal MRI. Through self-supervised pre-training, BSTNet aims to achieve model generalization across multi-timepoint and dual-timepoint scenarios while capturing dynamic tumor changes during NAT. In a multicenter cohort of 1339 patients, BSTNet demonstrated robust performance: area under the receiver operating characteristic curve (AUC) of 0.882 in internal validation (from Center 1), and 0.857 and 0.854 in external validation (Centers 2 and 3, respectively). In all three validation cohorts, subgroup analyses demonstrated consistent performance across molecular subtypes, with AUCs ranging from 0.827 to 0.886 for hormone receptor status and 0.818-0.895 for human epidermal growth factor receptor 2 status. The model maintained a stable performance across varying interim MRI timings in the external validation cohorts, with AUCs of 0.841-0.893 in Center 2 and 0.792-0.970 in Center 3. Notably, BSTNet effectively identified patients with non-pCR (specificity: 86.4%, 74.5%, and 85.1% for internal validation, Center 2, and Center 3, respectively). In conclusion, BSTNet provides a robust and generalizable deep learning framework for early pCR prediction. Its ability to effectively interpret variable longitudinal MRI data offers a powerful and practical tool to guide adaptive treatment planning in diverse clinical settings.
© 2026. The Author(s).