Explaining the Rationale of Deep Learning Glaucoma Decisions with Adversarial Examples.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning models (DLMs) are increasingly used in medical diagnostics, but their
• black box
• nature can limit clinician trust.
• Understanding the rationale behind DLM decisions is crucial for clinical adoption, especially in complex areas like glaucoma detection.

Purpose of the Study:

• To evaluate adversarial explanation (AE) as a method for elucidating the decision-making process of DLMs in identifying glaucoma and related findings.
• To compare the explainability of AE against a conventional heatmap-based method, gradient-weighted class activation mapping (GradCAM).

Main Methods:

• Deep learning models were trained on 6,430 retinal fundus images for referable glaucoma (RG), increased cup-to-disc ratio (ICDR), disc rim narrowing (DRN), and retinal nerve fiber layer defect (RNFLD).
• Adversarial examples (AEs) and GradCAM heatmaps were generated for 400 patient eyes to explain DLM decisions.
• Glaucoma specialists rated the location and rationale explainability of both AE and GradCAM methods via surveys.

Main Results:

• DLMs achieved high performance, with AUCs ranging from 0.79 to 0.99 for the tested conditions.
• Adversarial examples demonstrated valid clinical feature changes relevant to glaucoma.
• AEs significantly outperformed GradCAM in both location explainability (3.94 vs. 2.55) and rationale explainability (3.97 vs. 2.10) based on specialist ratings.

Conclusions:

• Adversarial explanation offers superior interpretability for deep learning models in ophthalmology compared to GradCAM.
• This enhanced explainability can foster greater clinician confidence and facilitate the integration of AI tools into clinical practice for glaucoma diagnosis.