GRACE-ViT: Grouped Recalibration With Adaptive Contextual Emphasis for Breast Cancer Neoadjuvant Chemotherapy Response Prediction.
Authors
Affiliations (5)
Affiliations (5)
- Bioengineering Department, J.B. Speed School of Engineering, University of Louisville, Louisville, KY, USA.
- Department of Computer Sciences, Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.
- Department of Information Sciences, Princess Nourah bint Abdulrahman University, Saudi Arabia.
- Pathology Department, Mansoura University, Mansoura, Egypt.
- Department of Radiology, University of Louisville, Louisville, KY, USA.
Abstract
IntroductionAccurate prediction of neoadjuvant chemotherapy (NAC) response in breast cancer is important for treatment planning and individualized therapy selection. Breast MRI contains subtle patterns related to treatment response. These patterns vary across image regions and may be difficult for standard deep learning models to capture. In this retrospective clinical prediction model development and validation study, we propose GRACE-ViT, a Vision Transformer-based framework for three-class NAC response prediction from pre-treatment breast MRI. The three classes are partial response, complete response, and stable disease.MethodsGRACE-ViT uses a pretrained ViT backbone with a lightweight token recalibration module called Grouped Recalibration with Adaptive Contextual Emphasis (GRACE). The GRACE module refines patch-token features using dual-mode token scoring, a spatial coherence prior, and cross-group feature integration. This helps the model focus on important breast regions while keeping global image context. The model was evaluated on 736 axial T1-weighted breast MRI images. It was compared with convolutional, transformer-based, and medical-imaging models under the same preprocessing, training, validation, and testing protocol. Performance was measured using accuracy, mean F1-score, precision, and recall where bootstrap resampling was used to estimate 95% confidence intervals.ResultsGRACE-ViT reached a mean accuracy of 95.50% (95% CI: 90.99-99.10%) and a mean F1-score of 95.47% (95% CI: 91.10-99.06%). It outperformed the strongest baseline with only a small increase in computational cost. The per-class results were balanced across the three response categories, suggesting that the model did not favor one clinical outcome over another. Visual results showed that the model mainly focused on clinically relevant breast regions rather than background areas.ConclusionsThese results show that targeted token recalibration can improve ViT-based breast MRI analysis without greatly increasing model complexity, leading GRACE-ViT to provide an efficient and interpretable imaging-based approach for predicting NAC response and supporting individualized treatment decisions in breast cancer care.