A Novel Convolutional Neural Network for Explainable Diabetic Retinopathy Detection and Grade Identification

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy is a leading cause of global blindness.
• Early diagnosis is crucial for effective intervention.
• Automated detection methods can aid clinical assessment.

Purpose of the Study:

• To develop an explainable deep learning method for automatic diabetic retinopathy severity identification.
• To evaluate the performance of various convolutional neural network architectures, including a novel FGNet.
• To provide visual insights into model predictions for clinical support.

Main Methods:

• Utilized deep learning with multiple CNN architectures (VGG16, StandardCNN, ResNet, CustomCNN, EfficientNet, MobileNet, FGNet).
• Developed and trained a novel architecture, FGNet, specifically for diabetic retinopathy detection.
• Employed Gradient-weighted Class Activation Mapping (Grad-CAM) for model explainability.

Main Results:

• The proposed FGNet achieved an accuracy of 0.75 after 10 epochs and 0.71 after 20 epochs.
• Explainability techniques provided visual insights into the model's decision-making process.
• The method demonstrated potential for supporting early clinical assessment.

Conclusions:

• The developed explainable AI method shows potential for accurate diabetic retinopathy severity detection.
• FGNet offers a promising approach for automated analysis of retinal images.
• Explainable AI in this context can enhance clinical decision-making and patient outcomes.