A Deep Learning Model to Detect Center-Involved Diabetic Macular Edema from Fundus Images

Abstract

Purpose : To develop a deep learning (DL) model to detect center-involved diabetic macular edema (ci-DME) from color fundus photographs (CFPs), assess its performance on multiple datasets and compare it with human graders, against a ground truth obtained from optical coherence tomography (OCT) scans.

Methods : We developed a DL model to detect ci-DME using a retrospective development dataset of 7341 CFPs from 4682 patients as input and the corresponding ci-DME grades based on retinal thicknesses from OCT macular scans as labels. Three validation datasets from three different countries were assessed by human graders for the presence of ci-DME, using the proxy criteria of presence of hard exudates within one optic disc diameter of the center of the macula (Thailand: 1033 CFPs, graded by 3 retina specialists; US: 1976 CFPs, graded by 1 certified grader; and UK: 1602 CFPs, graded by 1 certified grader). ci-DME from OCT was defined as eyes with ≥ 250 μm center point thickness in the Thailand dataset and ≥ 300 μm central subfield thickness in the US and UK datasets. We evaluated the model on these datasets and compared its performance to human graders, using ci-DME ground truth derived from OCT.

Results : As shown in Figure 1, on Thailand, US, and UK validation sets, the model achieved area under the curves (AUCs) of 0.880 (95% confidence interval (CI): [0.86,0.90]), 0.875 (95% CI: [0.83,0.92]) and 0.819 (95% CI: [0.78,0.85]) respectively. When matched with the sensitivities of approximately 85%, 40% and 93% with corresponding specificities of 45%, 94% and 32% of human graders, the model had specificities of 73%, 97% and 44% respectively (p < 0.01 for each comparison).

Conclusions : Our DL model achieved good performance in detecting ci-DME from CFPs in three validation sets from three different countries, with higher specificities when matching the sensitivities of human graders. Such a model could potentially be of clinical value in reducing both false positive referrals and false negative diagnoses for ci-DME in DR screening programs worldwide.

This is a 2020 ARVO Annual Meeting abstract.

 

Figure 1: Receiver operating characteristics (ROC) curves for the ci-DME model evaluated on Thailand, US and UK validation datasets, together with the AUCs and graders' performance.