Interpretable Diabetic Retinopathy Diagnosis Based on Biomarker Activation Map.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning classifiers offer high accuracy in diagnosing diabetic retinopathy (DR) using optical coherence tomography (OCT) and OCT angiography (OCTA).
• The complex hidden layers in these models, while powerful, limit interpretability, hindering clinical trust and verification.
• Understanding the decision-making process of AI in DR diagnosis is crucial for clinical adoption.

Purpose of the Study:

• To introduce a novel biomarker activation map (BAM) framework for enhancing the interpretability of deep learning classifiers for DR diagnosis.
• To enable clinicians to verify and understand the decision-making processes of AI models in identifying DR.

Main Methods:

• Developed a BAM framework utilizing generative adversarial learning with two U-shaped generators.
• Trained a DR classifier on 456 macular scans graded for DR.
• Generated BAMs by contrasting input scans with outputs from a generator trained to alter classification, highlighting key biomarkers.

Main Results:

• The generated BAMs successfully highlighted known pathological features of DR, such as nonperfusion areas and retinal fluid.
• The framework demonstrated the ability to visualize the specific features driving the classifier's decisions.

Conclusions:

• The BAM framework provides a method for creating interpretable deep learning models for DR diagnosis.
• An interpretable classifier can improve clinician confidence and facilitate the verification of automated DR diagnoses.