DR.DIVYA RAO

DR. ANAND VINEKAR, FLORIAN M. SAVOY, JUN HUP LIM, JUN LEONG HOE

Abstract

Purpose: Artificial Intelligence (AI) can improve diagnostic consistencies and help scale up ROP screening. We developed an AI algorithm to automatically detect the presence of ROP in retinal images.

Methods: We trained a deep learning algorithm with 28,975 temporal images from a tele-ROP screening program (Retcam cameras). The classifier is trained to segregate no ROP images and images with ROP of stage 2 and above. The labels consist of per-eye diagnoses by trained ROP image graders during screening. Those requiring treatment had a confirmatory diagnosis from clinical exam.

Results: Of the 3291 images in test set, 33.4% showed ROP (stage 2 and above). Sensitivity was 92.26% (95% CI: 90.52% to 93.77%) with 85 positive images misclassified. Specificity was 87.73% (95% CI:86.29% to 89.08%), with 269 images with no ROP misclassified.

Conclusions: The tool shows promising performance. A prospective clinical validation in a real-world tele-ROP platform is under consideration.