Enhancing AI-based diabetic retinopathy diagnosis through universal cross-camera image adaptation.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness globally.
• Current AI models for DR detection often struggle with generalisability across different fundus camera systems.
• Variability in image acquisition and style can significantly impact AI diagnostic performance.

Purpose of the Study:

• To evaluate a deep learning-based style adaptation strategy for improving AI diagnostic accuracy in DR detection.
• To assess the cross-camera generalisability of AI models after style adaptation.
• To investigate the impact of style adaptation on reducing false positives and maintaining diagnostic performance.

Main Methods:

• Prospective diagnostic study involving patients aged 50+ at a tertiary eye hospital in India.
• Paired retinal images captured using Optain Resolve and Topcon NW400 fundus cameras.
• Application of Style-Consistent Retinal Image Transformation Network (SCR-Net) for style alignment.
• Evaluation of an InceptionNeXt-T based AI model under different training and testing scenarios, including style-adapted images.

Main Results:

• The mixed training/testing approach with style adaptation achieved the highest diagnostic accuracy (79.2%) for Optain images.
• Style adaptation preserved critical diagnostic features and image quality (PSNR: 29.35, SSIM: 0.847).
• Reduced false positives in Optain images and maintained robust performance for Topcon images, addressing cross-camera variability.

Conclusions:

• Style adaptation using SCR-Net enhances AI consistency and generalisability for DR detection.
• This approach reduces false positives and maintains robust performance across diverse camera systems.
• Potential to democratise early DR diagnosis, though further validation across diverse settings is needed.