A Study on the Interpretability of Diabetic Retinopathy Diagnostic Models

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Diabetic retinopathy detection

Background:

• Diabetic retinopathy is a leading cause of blindness globally.
• Accurate classification of diabetic retinopathy is crucial for timely treatment.
• Deep learning models show promise but require interpretability for clinical trust.

Purpose of the Study:

• To assess the interpretability of deep learning models for diabetic retinopathy classification.
• To compare the performance of seven interpretability methods across four distinct model architectures.
• To provide insights into how model architecture impacts classification interpretability.

Main Methods:

• Applied seven interpretability methods (Gradient, SmoothGrad, Integrated Gradients, SHAP, DeepLIFT, Grad-CAM++, ScoreCAM) to VGG, ResNet, DenseNet, and EfficientNet models.
• Utilized saliency map visualization, perturbation curve analysis, and trend correlation analysis.
• Quantitatively evaluated interpretability using saliency map entropy, AOPC score, Recall, and Dice coefficient.

Main Results:

• Model architecture significantly affects interpretability quality.
• Simpler architectures (e.g., VGG) with clearer feature extraction paths demonstrated superior interpretability.
• Deeper or lightweight architectures presented certain interpretability limitations.

Conclusions:

• The choice of deep learning architecture is a critical factor in achieving reliable model interpretability for diabetic retinopathy classification.
• Future research should consider architectural design for enhanced explainability in medical AI.
• Understanding model interpretability is key to clinical adoption of AI in ophthalmology.