DR.DIVYA RAO

FLORIAN M. SAVOY, JUN KAI TOH

Abstract

Purpose: We developed an automated screening tool for AMD using DL on fundus images. It is deployable offline on a portable camera and runs in seconds.

Methods: Training relied on 128,015 images (47% referable AMD) from the AREDS study (image grading by a central reading center) and 598 images (26% referable) from Asian eyes using the target device (consensus grading by two vitreo-retinal specialists). The model indicates referable AMD (intermediate and advanced). Validation set comprises of 334 images (34% referable) and test set comprises of 332 images (33% referable), both from target device.

Results: Sensitivity was 87.7%, specificity 85.0% and AUC 0.92 in detecting referable AMD on the validation set. On test set, sensitivity was 78.9%, specificity 84.8% and AUC 0.90

Conclusion: The model shows promising results, despite relying on a training set predominantly captured by traditional cameras and different population. It can make AMD screening accessible, affordable and effective.