[Paper Review] Interpretable Medical Imagery Diagnosis with Self-Attentive Transformers: A Review of Explainable AI for Health Care

Tin Lai|arXiv (Cornell University)|Sep 1, 2023

Explainable Artificial Intelligence (XAI)Computer Science3 citations

TL;DR

This paper reviews self-attentive Vision Transformers (ViT) for interpretable medical image diagnosis, proposing XAI techniques to decode ViT decision-making processes. By leveraging attention maps and saliency visualization, the approach enhances transparency in AI-driven medical diagnostics, enabling clinicians to trust and validate model predictions in clinical settings.

ABSTRACT

Recent advancements in artificial intelligence (AI) have facilitated its widespread adoption in primary medical services, addressing the demand-supply imbalance in healthcare. Vision Transformers (ViT) have emerged as state-of-the-art computer vision models, benefiting from self-attention modules. However, compared to traditional machine-learning approaches, deep-learning models are complex and are often treated as a "black box" that can cause uncertainty regarding how they operate. Explainable Artificial Intelligence (XAI) refers to methods that explain and interpret machine learning models' inner workings and how they come to decisions, which is especially important in the medical domain to guide the healthcare decision-making process. This review summarises recent ViT advancements and interpretative approaches to understanding the decision-making process of ViT, enabling transparency in medical diagnosis applications.

Motivation & Objective

• To address the lack of interpretability in deep learning models used for medical imaging, particularly Vision Transformers (ViT).
• To review recent advancements in ViT architectures and their integration with Explainable AI (XAI) methods.
• To analyze how attention mechanisms in ViTs can be leveraged to generate human-understandable explanations for medical image diagnoses.
• To evaluate the effectiveness of XAI techniques in improving transparency and clinical trust in AI-driven diagnostic systems.
• To provide a comprehensive overview of current challenges and future directions in interpretable AI for healthcare.

Proposed method

• Utilizes Vision Transformers (ViT) as the core deep learning architecture for medical image analysis, leveraging multi-head self-attention mechanisms.
• Applies attention map visualization to highlight regions of interest in medical images that most influenced the model's predictions.
• Employs saliency-based explanation methods such as Grad-CAM and Integrated Gradients to localize decision-relevant features in input images.
• Integrates class activation mapping techniques adapted for ViT to generate class-specific attention heatmaps.
• Reviews and compares various XAI techniques tailored for ViT, focusing on their applicability and reliability in clinical imaging tasks.
• Analyzes the fidelity and robustness of attention-based explanations across diverse medical imaging modalities (e.g., X-ray, MRI, CT).

Experimental results

Research questions

• RQ1 How can Vision Transformers be made interpretable for medical image diagnosis using attention-based explanations?
• RQ2 What are the most effective XAI techniques for visualizing decision-making in ViT models applied to medical imaging?
• RQ3 How do attention maps in ViTs compare to traditional CNN-based saliency maps in terms of clinical interpretability?
• RQ4 To what extent do XAI methods improve clinician trust and understanding of AI-generated diagnoses?
• RQ5 What are the limitations and challenges of applying XAI to ViT-based medical diagnostic systems?

Key findings

• Self-attention mechanisms in Vision Transformers provide intrinsic interpretability by highlighting relevant image regions through attention weights.
• Attention-based explanations in ViTs show improved localization of pathologies compared to traditional CNN-based methods in multiple medical imaging tasks.
• XAI techniques such as Grad-CAM and Integrated Gradients effectively generate class-specific saliency maps that align with radiological findings.
• The fidelity of attention maps in ViTs is higher than in convolutional models, particularly in complex anatomical structures.
• Clinicians report increased trust in AI predictions when provided with attention-based visual explanations.
• Despite progress, challenges remain in standardizing XAI evaluation and ensuring robustness across diverse medical imaging datasets.