Quantifying interpretation reproducibility in Vision Transformer models with TAVAC

Area of Science:

• Artificial Intelligence
• Digital Pathology
• Medical Imaging Analysis

Background:

• Deep learning, particularly Vision Transformer (ViT) models, shows promise for extracting diagnostic features from biomedical images, outperforming traditional CNNs in spatial relationship capture and interpretability.
• However, limited annotated datasets can lead to ViT model overfitting, resulting in inaccurate predictions due to noise, hindering clinical application.
• Ensuring the reliability and reproducibility of AI model interpretations is crucial for both clinical diagnostics and basic scientific research.

Purpose of the Study:

• To introduce Training Attention and Validation Attention Consistency (TAVAC), a novel metric for evaluating Vision Transformer (ViT) model overfitting in biomedical image analysis.
• To quantify the reproducibility of interpretations generated by ViT models, ensuring reliable feature extraction.
• To differentiate between on-target and off-target attention mechanisms within ViT models.

Main Methods:

• TAVAC was developed by comparing high-attention regions between training and testing phases of ViT models.
• The metric was validated on four public image classification datasets and two independent breast cancer histological image datasets.
• TAVAC's ability to distinguish between on-target and off-target attentions and measure interpretation generalization at a cellular level was assessed.

Main Results:

• Overfitted ViT models consistently exhibited significantly lower TAVAC scores compared to well-generalized models.
• TAVAC effectively distinguished between relevant (on-target) and irrelevant (off-target) attention patterns within the models.
• The metric demonstrated capability in measuring interpretation generalization at a fine-grained cellular level, applicable to both biomedical and general images.

Conclusions:

• TAVAC is a robust metric for evaluating ViT model overfitting and enhancing interpretative reproducibility in digital pathology and beyond.
• The metric aids in identifying unreliable predictions stemming from overfitting, thereby improving the trustworthiness of AI in medical diagnostics.
• TAVAC's application extends to basic research, facilitating the discovery of critical spatial patterns and cellular structures through reliable interpretation of imaging data.