A New Deep Learning Algorithm with Activation Mapping for Diabetic Retinopathy: Backtesting after 10 Years of...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Early detection and accurate diagnosis of DR are crucial for effective management.
• Current diagnostic methods rely on expert human interpretation of fundus images.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for detecting diabetic retinopathy from fundus images.
• To compare the performance of the AI algorithm against human diagnoses by family doctors (FDs).
• To assess the potential of AI as a diagnostic aid in DR screening.

Main Methods:

• Development of a ResNet-50 deep learning model with enhanced features (double resolution, Squeeze-Excitation blocks).
• Model pre-trained on ImageNet and trained for 50 epochs using the Adam optimizer.
• Utilized a large dataset of approximately 500,000 fundus images classified by clinicians.

Main Results:

• The AI algorithm achieved over 95% detection rate for cases worse than mild DR.
• The AI demonstrated 70% fewer misclassifications of healthy cases compared to FDs.
• The AI identified DR signs in 7.9% of patients before detection by FDs.

Conclusions:

• The developed AI algorithm demonstrates performance comparable to or exceeding that of family doctors in DR detection.
• AI-based tools show potential as valuable aids for DR diagnosis, improving early detection rates.
• The study suggests integrating AI as a diagnostic support tool rather than a standalone diagnostic solution.