Applicability study of AI attribution methods for ophthalmic image classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) and retinal fluid accumulation are major diabetes complications detectable via optical coherence tomography (OCT).
• Deep learning models show high accuracy in diagnosing these conditions but lack clinical trust due to poor interpretability.
• Attribution methods, successful in natural image analysis, are underexplored for medical imaging interpretability.

Purpose of the Study:

• To evaluate the effectiveness of three attribution methods (DeepLIFT, AGI, AttEXplore) in explaining deep learning model predictions for DR and fluid detection using OCT images.
• To assess the clinical relevance of highlighted regions identified by these methods using quantitative and qualitative measures.
• To determine the most suitable attribution method for improving transparency in AI-assisted ophthalmic diagnostics.

Main Methods:

• A VGG16 deep learning model was trained for DR classification and fluid detection using widefield OCTA en face and OCT B-scan images.
• Three attribution methods—DeepLIFT, AGI, and AttEXplore—were employed to generate visual explanations for the model's predictions.
• Quantitative metrics (insertion/deletion scores) and qualitative heatmap analysis were used to assess the clinical relevance of the attributions.

Main Results:

• The VGG16 model achieved high accuracy (94% for DR, 98% for fluid detection).
• Attribution methods produced varied qualitative results, influenced by their underlying assumptions and hyperparameter sensitivity.
• AttEXplore demonstrated superior performance in highlighting clinically meaningful structures in pathological cases compared to AGI and DeepLIFT, despite similar quantitative scores between AGI and AttEXplore.

Conclusions:

• While attribution methods can aid in explaining AI predictions for ophthalmic conditions, their results require careful interpretation alongside clinical expertise.
• AttEXplore shows potential as a preferred method for visualizing clinically relevant features in DR and fluid detection tasks.
• Further research is needed to establish attribution methods as reliable proxies for clinical relevance in AI-assisted medical diagnostics.