Using a Deep Learning-based OCT Image Analysis for Diabetic Macular Edema Identification to Support Triage Screening: a Prospective Study

Abstract

Purpose : Recent retrospective analysis has already demonstrated diabetic macular edema (DME) can be identified accurately from optical coherence tomography (OCT) images via deep learning (DL). However, data on how artificial intelligence (AI) improves current clinical workflow and patient outcomes in the healthcare system is scarce. Here, we prospectively evaluated the performance of DL algorithms that can assess image quality and identify DME from volumetric macular OCT scans (i.e., AI-OCT) for screening DME in a triage unit of a tertiary eye hospital.

Methods : Our AI-OCT system is a customized AI-powered infrastructure that has 2 core DL algorithms based on our previously published work: 1) gradability assessment for macular scans (gradable/ungradable scan) and 2) classification of DME (DME/no DME). Under the infrastructure, a browser/server system was designed and developed for real-time OCT image extraction, input data configuration (e.g., subject ID and exam date), image upload, and results output analysis via a graphics processing unit server. Only gradable scans would then be processed by the DME classification module. Subjects with diabetes mellitus (DM) referred to the triage unit were recruited. Participants underwent Cirrus OCT scans (Carl Zeiss Meditec, Dublin, CA), which were then analyzed by the AI-OCT system in real-time. Output analysis outcomes were automatically generated in a report. The performance of the AI-OCT system was measured against reference standards, with image gradability provided by well-trained graders and DME diagnosis provided by retina specialists.

Results : A total of 1200 eyes from 603 participants with DM were recruited. The performance of the image quality assessment module was evaluated, with a sensitivity (SEN) of 99% (95% CI 98%-100%), specificity (SPE) of 80% (70%-87%), and accuracy (ACC) of 98% (96%-98%). Among 1106 eyes with gradable scans (92%) and then processed by the DME classification module, the AI-OCT system had a SEN of 99% (93%-100%), SPE of 91% (89%-93%), and ACC of 92% (90%-93%) for DME detection.

Conclusions : In this prospective study, our AI-OCT system showed excellent performance for DME triage screening, further providing data on evidence-based healthcare implementation to move AI from “concept” to “real-world clinic.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1. Study flow diagram