GRACE-ViT: Grouped Recalibration With Adaptive Contextual Emphasis for Breast Cancer Neoadjuvant Chemotherapy...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate prediction of neoadjuvant chemotherapy (NAC) response in breast cancer is crucial for tailoring treatment plans.
• Breast MRI reveals subtle patterns indicative of treatment response, but these are challenging for standard deep learning models.
• Individualized therapy selection necessitates reliable methods for predicting NAC outcomes.

Purpose of the Study:

• To develop and validate GRACE-ViT, a Vision Transformer-based framework for predicting three-class NAC response (partial response, complete response, stable disease) from pre-treatment breast MRI.
• To enhance the analysis of subtle image patterns in breast MRI for improved NAC response prediction.
• To provide an efficient and interpretable imaging-based tool for supporting individualized breast cancer treatment decisions.

Main Methods:

• Developed GRACE-ViT, a Vision Transformer framework incorporating a Grouped Recalibration with Adaptive Contextual Emphasis (GRACE) module for refining image patch-token features.
• Employed dual-mode token scoring, spatial coherence prior, and cross-group feature integration within the GRACE module to focus on relevant breast regions while maintaining global context.
• Evaluated the model on 736 axial T1-weighted breast MRI images, comparing its performance against baseline models using accuracy, F1-score, precision, and recall with bootstrap resampling for confidence intervals.

Main Results:

• GRACE-ViT achieved 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%).
• The model demonstrated superior performance compared to the strongest baseline models with only a marginal increase in computational cost.
• Per-class results were balanced, indicating no bias towards specific clinical outcomes, and visual analysis confirmed the model's focus on clinically relevant breast regions.

Conclusions:

• Targeted token recalibration in Vision Transformer models can significantly improve breast MRI analysis for NAC response prediction without substantial increases in complexity.
• GRACE-ViT offers an efficient, interpretable, and accurate imaging-based approach for predicting neoadjuvant chemotherapy response in breast cancer.
• The findings support the potential of GRACE-ViT in facilitating individualized treatment strategies and improving patient care in breast cancer management.