Enhancing Breast Cancer Diagnosis With Multi-Resolution Vision Transformers and Robust Decision-Making

This study aims to improve breast cancer (BC) diagnosis through a novel multi-resolution Vision Transformer (ViT)-based framework with ensemble decision-making, addressing limitations in traditional single-magnification models. The proposed framework uses multiscale feature extraction at three magnification levels (L0, L1, L2 or 16x, 4x, 2x) to capture both fine-grained and high-level tumor features. A stacking ensemble method combines predictions from ViT models trained at these levels, improving classification robustness. Postprocessing techniques, including region-growing and fast-marching level set algorithms, refine whole-slide image (WSI) prediction and postprocessing quality. Performance was evaluated via metrics such as precision, recall, the F1 score, accuracy, and specificity across 50 trials with perturbed conditions. The framework achieved a top accuracy of 97.08%, with precision and recall above 94%. The suggested stack configuration outperformed individual models and other stacking configurations, demonstrating balanced performance with minimal variability. Statistical analysis highlighted the reliability and consistency of the framework under perturbed conditions. The multi-resolution ViT-based framework significantly improves BC classification by integrating multiscale analysis and ensemble decision-making. Its high accuracy and robustness make it a valuable tool for reducing interobserver variability in digital pathology workflows.