A latent diffusion approach to visual attribution in medical imaging.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computer Vision in Radiology
• Explainable AI for Healthcare

Background:

• Standard machine vision for medical imaging often detects disease but lacks interpretability for clinicians.
• Visual attribution aims to identify diagnostically relevant image components, enhancing clinical understanding.
• Existing methods may not fully bridge the gap between AI detection and clinical interpretability.

Purpose of the Study:

• To develop a novel generative visual attribution technique for medical imaging.
• To enhance the interpretability and explainability of AI models in clinical radiology.
• To generate normal counterparts of abnormal medical images for comparative analysis.

Main Methods:

• Leveraging latent diffusion models (LDMs) combined with domain-specific large language models (LLMs).
• Generating normal reference images by controlling the image generative process using image priors and conditioning mechanisms.
• Utilizing natural language text prompts derived from medical science and radiology for conditioning the generative models.
• Employing a discrepancy analysis between abnormal and generated normal images to create attribution maps.

Main Results:

• Quantitative evaluation on the COVID-19 Radiography Database using Frechet Inception Distance (FID), Structural Similarity (SSIM), and Multi-Scale Structural Similarity (MS-SSIM) metrics.
• Demonstrated effectiveness in generating realistic normal counterparts of abnormal chest X-rays.
• The developed system shows latent capabilities, including zero-shot localized disease induction, validated on the cheXpert dataset.

Conclusions:

• The proposed generative visual attribution technique effectively highlights diagnostically relevant regions in medical images.
• The integration of LDMs and LLMs offers a powerful approach for interpretable medical AI.
• The method shows promise for improving clinical decision support by providing clear visual explanations for AI findings.