Explainable end-to-end deep learning for diabetic retinopathy detection across multiple datasets.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of visual impairment globally.
• Manual analysis of retinal fundus images for DR diagnosis is time-consuming.
• Automated systems can improve DR detection efficiency and reduce healthcare costs.

Purpose of the Study:

• To develop a deep learning algorithm for automated DR detection in retinal fundus images.
• To assess the algorithm's performance across multiple public datasets.
• To implement an explainability method for visualizing DR signs identified by the model.

Main Methods:

• Fine-tuning a pretrained deep convolutional neural network for DR classification.
• Training the model on over 90,000 images from nine public datasets.
• Utilizing a cosine annealing learning rate strategy with warm-up for improved training.
• Employing gradient-weighted class activation mapping for model explainability.

Main Results:

• Achieved high classification performance with an AUC of 0.986 on the EyePACS dataset.
• Demonstrated strong AUC values (0.957-0.990) across eight additional datasets.
• The explainability model successfully visualized DR signs detected by the deep learning network.

Conclusions:

• The developed deep learning approach achieves state-of-the-art performance in DR detection.
• The algorithm robustly classifies fundus images and outperforms previous methods using public datasets.
• The explainability feature enhances trust and understanding of the automated DR detection process.