Optimizing Vision Transformers for Histopathology: Pretraining and Normalization in Breast Cancer Classification.

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Artificial Intelligence in Oncology

Background:

• Accurate breast cancer classification from histology images is crucial for effective treatment.
• Vision Transformers (ViTs) show promise for image classification tasks.
• Optimizing ViT performance for histopathology requires careful consideration of training strategies.

Purpose of the Study:

• To develop and evaluate a self-attention Vision Transformer model for breast cancer classification in histology images.
• To investigate the impact of various training strategies on model performance.
• To assess the generalization capabilities of the proposed model across different datasets.

Main Methods:

• A self-attention Vision Transformer model was developed.
• Extensive evaluation of training strategies including pretraining, data augmentation, and patch configurations.
• Models were trained and validated on the BACH dataset and tested on BRACS and AIDPATH datasets.

Main Results:

• The ImageNet-pretrained ViT achieved accuracies of 0.91 (BACH), 0.74 (BRACS), and 0.92 (AIDPATH).
• Geometric and color data augmentation techniques demonstrated increased effectiveness.
• Domain-specific pretraining showed potential but did not yet offer clear advantages over general pretraining.

Conclusions:

• Pretraining on large-scale general datasets like ImageNet is beneficial for histology image classification.
• The developed Vision Transformer model demonstrates strong performance and generalization capabilities.
• Further research into domain-specific pretraining could enhance histopathology image analysis.