Related Experiment Video
Updated: May 6, 2026
Using Retinal Imaging to Study Dementia
Published on: November 6, 2017
Enhancing AI-based diabetic retinopathy diagnosis through universal cross-camera image adaptation
Sanil Joseph1,2,3, Xiaotian Chen4, Chi Liu5
1Ophthalmic Epidemiology, Centre for Eye Research Australia, East Melbourne, Victoria, Australia sanil@aravind.org Zongyuan.ge@monash.edu mingguang.he@polyu.edu.hk.
Objective:
To evaluate the effectiveness of a deep learning-based style adaptation strategy in improving the diagnostic accuracy and cross-camera generalisability of artificial intelligence (AI) for detecting diabetic retinopathy (DR).
Methods And Analysis:
This diagnostic study involved prospective recruitment of patients aged 50 years and older attending the outpatient clinic at a tertiary eye hospital in Southern India, between 14 June and 5 August 2022. Paired macula-centred retinal images were captured using two fundus cameras: Optain Resolve (portable, automated) and Topcon NW400 (static, manual). A style adaptation model, the Style-Consistent Retinal Image Transformation Network (SCR-Net), was applied to align image styles across cameras. The AI-based DR detection model, developed using the InceptionNeXt-T architecture, was trained on images from the EyePACS data set and evaluated under three scenarios: (1) training and testing on original images (2) training and testing on SCR-Net-adapted images; and (3) training on a combined (original+adapted) data set and testing on adapted images. Diagnostic accuracy and preservation of image quality were evaluated.
Results:
The mixed training/testing approach (scenario 3) achieved the highest diagnostic accuracy for Optain images at 79.2% (95% CI 75.9% to 82.6%) with a Cohen's kappa of 0.893 (95% CI 0.867 to 0.917). Adapted images preserved critical diagnostic features (peak signal-to-noise ratio, 29.35; structural similarity index measure, 0.847). Style adaptation reduced false positives in Optain images while maintaining robust diagnostic performance for Topcon images, effectively addressing cross-camera variability.
Conclusion:
Style adaptation using SCR-Net enhances the consistency and generalisability of AI-based DR detection systems by reducing false positives and maintaining robust performance across camera systems. This approach has the potential to democratise access to early DR diagnosis in underserved regions. This study was conducted at a single centre using a limited set of fundus cameras, which may affect the generalisability. Nonetheless, further validation across diverse imaging systems and clinical settings is warranted to support broader applicability.
