MGRW-Transformer: Multigranularity Random Walk Transformer Model for Interpretable Learning.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Computer Vision

Background:

• Deep learning models excel at image recognition but often function as "black boxes," lacking semantic explanations.
• Vision Transformers (ViT) offer improved interpretability via self-attention mechanisms.
• Challenges remain in applying deep learning to medical images due to variable lesion sizes and locations, hindering accurate and explainable conclusions.

Purpose of the Study:

• To develop an interpretable deep learning model for medical image recognition.
• To address the limitations of current models in providing semantic explanations for their decisions.
• To enhance the application of AI in medical diagnostics by offering visual insights into classification processes.

Main Methods:

• Proposed a multigranularity random walk transformer (MGRW-Transformer) model integrating attention mechanisms with a random walk approach.
• Divided medical images into subimage blocks, processed by a ViT module for classification.
• Fused the attention matrix with a multigranularity random walk module, constructing a graph where image blocks are nodes and attention guides the walk.

Main Results:

• The MGRW-Transformer model provides semantic interpretations and visualizations of the decision-making process.
• Experimental results demonstrate improved classification performance on medical images compared to existing methods.
• The model offers clear insights into how conclusions are reached, aiding medical professionals.

Conclusions:

• The MGRW-Transformer model successfully enhances interpretability in medical image recognition.
• The proposed method offers a valuable tool for medical professionals by providing explainable AI insights.
• This approach advances the integration of deep learning in clinical practice through enhanced transparency.