RetCond: A Conditional Diffusion Model for Self-Explanatory Multi-Class Fundus Image Classification

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Vision impairment is a global health issue, necessitating early retinal disease diagnosis.
• Manual interpretation of color fundus photography is time-consuming and variable.
• Current deep learning models for retinal disease detection often lack transparency.

Purpose of the Study:

• To introduce RetCond, a self-explanatory generative classifier for retinal disease detection.
• To provide accurate predictions with built-in, counterfactual visual explanations.
• To enhance trust and transparency in AI solutions for ophthalmology.

Main Methods:

• RetCond, a diffusion model repurposed as a classifier, generates counterfactual images for explanations.
• A diverse dataset of 19,565 retinal images across five disease conditions was curated.
• Performance was evaluated using standard classification metrics and qualitative/quantitative analysis of explanations.

Main Results:

• RetCond achieved 96.98% classification accuracy, comparable to state-of-the-art models.
• The model generated condition-specific counterfactual images, confirming its learned concepts.
• RetCond demonstrated self-explanatory properties, highlighting its decision-making process.

Conclusions:

• RetCond offers trustworthy AI in ophthalmology by matching discriminative performance with transparency.
• The generative classifier addresses limitations of post-hoc interpretability in deep learning.
• Self-explanatory AI models like RetCond are crucial for reliable automated retinal disease diagnosis.