Aravind Eye Hospital achieves 79.2% DR diagnostic accuracy with cross-camera AI style adaptation

AI-based diabetic retinopathy (DR) detection systems trained on images from one fundus camera type perform inconsistently when deployed with different camera hardware. Portable cameras used in low-resource settings produce images with different resolution, contrast, and colour characteristics, causing high false-positive rates and reducing clinical reliability.

Researchers deployed a Style-Consistent Retinal Image Transformation Network (SCR-Net) at AEH to standardise retinal image style across portable (Optain Resolve) and static (Topcon NW400) fundus cameras. The adaptation model does not require prior access to source-style images, making it universally applicable. A mixed training approach — combining original and SCR-Net-adapted images — was used to train an InceptionNeXt-T deep learning model for DR classification.

The mixed training/testing scenario achieved the highest accuracy of 79.2% (95% CI 75.9%–82.6%) with a Cohen's kappa of 0.893 for the portable Optain camera, up from 75.9% baseline without adaptation. Style adaptation significantly reduced false positives in portable camera images while maintaining stable performance for the standard Topcon camera. Adapted images preserved diagnostic quality with a PSNR of 29.35 and SSIM of 0.847.

• Universal style adaptation that requires no source-domain training data is viable and scalable for multi-camera DR screening deployments.
• Mixing original and adapted images during training outperforms training only on adapted images, preserving model robustness across camera types.
• Portable fundus cameras can achieve near-parity with clinical-grade cameras when paired with AI style normalisation, expanding DR screening access in underserved regions.