Abstract
Purpose :
AMD is one of the leading causes of blindness in the world. We developed and assessed an automated screening tool for AMD using deep learning (DL) on color fundus images. The model is deployable on-the-edge on the target device, a smartphone based portable fundus camera.
Methods :
We trained an AI binary classifier in two steps. We used two distinct datasets with macula centered images: 108,251 images (55% referable AMD) from the Age-related Eye Disease Study (AREDS), and 598 images (26% referable AMD) captured on Asian eyes using the target device. The model was trained to indicate the presence of referable AMD (intermediate and advanced AMD). We first trained a base model using AREDS only, with a validation set (A) of 990 images (45% referable AMD). We then finetuned this model using the target device dataset only. The finetuning validation set (B) comprised 334 images (34% referable AMD) and test set comprised 332 images (33% referable AMD), both from target device. The reference standard for AMD diagnosis on the AREDS dataset was fundus image grading by a central reading center. On the target dataset, it was the consensus image grading of two vitreo-retinal specialists
Results :
The DL algorithm had sensitivity of 94.7% (95% CI: 88.9% to 98.0%), specificity of 87.7% (95% CI: 82.7% to 91.8%) and AUC of 0.98 (95% CI: 0.95 to 1.00) in detecting referable AMD on the validation set B. On the test set, sensitivity was 86.2% (95% CI: 78.3% to 92.1%), specificity was 88.8% (95% CI: 83.9% to 92.6%) and AUC 0.95 (95% CI: 0.92 to 0.98). The model before finetuning had a validation set (A) sensitivity of 89.7% (95% CI: 86.5% to 92.4%), specificity of 93.6% (95% CI: 91.6% to 95.3%) and AUC of 0.97 (95% CI: 0.96 to 0.98).
Conclusions :
The DL algorithm shows promising results, despite the small size of the target device and population dataset. This is made possible by finetuning a model previously trained for the same task on a much larger dataset. The strategy proved helpful despite AREDS being captured by traditional desktop cameras on a different population. This on-the-edge AI deployable on a portable camera has potential to make AMD screening accessible, affordable, and effective.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.