Abstract PS2-09-01: Optimizing Transfer Learning for Early Prediction of Pathologic Complete Response in Breast Cancer: A Multicenter Study Using Pretreatment DCE-MRI

Abstract Background: Pathologic complete response (pCR) after neoadjuvant therapy (NAT) in breast cancer is associated with improved outcomes; however, pCR rate varies dramatically (e.g., from 8-65%) depending on molecular subtype and choice of therapy. While transfer learning from pretrained deep learning models has shown promise in advancing early prediction of pCR with DCE-MRI, the impact of the pretraining data domain remains underexplored, particularly across molecular subtypes. Objective: To evaluate the impact of general-domain versus domain-specific pretraining in transfer learning for pCR prediction from pretreatment DCE-MRI. Materials and Methods: We leveraged the multicenter MAMA-MIA dataset of pretreatment DCE-MRI from patients undergoing NAT (1357 patients; pCR: 424, non-pCR: 933) to evaluate ResNet-50 models pretrained on (1) ImageNet: a multi-source dataset of 1.28M natural images and thousands of labels, and (2) RadImageNet: a single-site dataset of 1.35M radiologic images and 185 labels. Both models were fine-tuned (top 10% layers) and evaluated for pCR prediction with pretreatment DCE-MRI, using identical hyperparameters and five-fold stratified cross-validation preserving center ratios. Predictive performance was assessed overall and across tumor subtypes: luminal A, luminal B, triple-negative, and HER2-enriched. Areas under the curve (AUCs) were compared using DeLong’s test. The combined performance of the best fine-tuned model and clinical features (HER2, HR status) was also assessed. Results: Transfer learning from ImageNet significantly outperformed RadImageNet in pCR prediction with pretreatment DCE-MRI overall (AUC: 0.63 vs. 0.59; p = 0.02). The performance of both fine-tuned models varied across different tumor subtypes, with statistically significant AUC gains observed for ImageNet relative to RadImageNet in luminal A (0.66 vs. 0.61; p = 0.01) and HER2-enriched tumors (0.68 vs. 0.53; p = 0.009). No significant differences between the two models were observed for luminal B or triple-negative tumors. Combination of the fine-tuned ImageNet model with clinical features (HER2, HR status) further enhanced pCR prediction, with the combined model achieving an AUC of 0.68. Conclusion: Based on our preliminary findings, ImageNet outperformed RadImageNet in transfer learning for predicting pCR from pretreatment breast DCE-MRI. This superior performance is likely due to the richer and more diverse features learned from the multisource general-domain images in ImageNet. Furthermore, the tumor subtype significantly influenced the performance of both models, and performance improvement with ImageNet. Our results provide valuable insights towards optimizing the use of transfer learning in predictive modeling for early prediction of pCR with DCE-MRI in breast cancer patients undergoing NAT. Citation Format: K. Bhalla, A. Sanchez, J. Luna, E. Podany, T. Ahmad, D. Bennett, A. Davis, A. Gastounioti. Optimizing Transfer Learning for Early Prediction of Pathologic Complete Response in Breast Cancer: A Multicenter Study Using Pretreatment DCE-MRI abstract. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS2-09-01.