ExplAIn: Explanatory artificial intelligence for diabetic retinopathy diagnosis

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Artificial Intelligence (AI) shows promise in medical decision support.
• The "black-box" nature of current AI hinders widespread clinical adoption.
• Diabetic Retinopathy (DR) classification from Color Fundus Photography (CFP) is a critical diagnostic task.

Purpose of the Study:

• To introduce an eXplanatory Artificial Intelligence (XAI) model, ExplAIn, for DR severity classification.
• To achieve performance comparable to black-box AI models while providing explainability.
• To facilitate the deployment of AI in clinical settings by addressing the transparency issue.

Main Methods:

• ExplAIn is an end-to-end trained XAI algorithm using only image supervision.
• The model learns to segment and categorize lesions, with classification derived from these segmentations.
• Self-supervision is employed for foreground/background separation to enhance lesion localization.

Main Results:

• ExplAIn achieves high performance in classifying DR severity using CFP.
• The model demonstrates the ability to segment and categorize relevant lesions within images.
• Explanations are generated through image-based lesion localization and categorization.

Conclusions:

• The ExplAIn framework offers both high classification performance and explainability for DR detection.
• This XAI approach can simplify AI-driven diagnoses through visual and textual outputs.
• The ExplAIn model is expected to accelerate the integration of AI in ophthalmology and medical imaging.