Abstract
Purpose :
Diabetic macular edema (DME) is a leading cause of vision loss in patients with diabetes. DME is frequently accompanied by other retinal abnormalities on two-dimensional optical coherence tomography (2D OCT). We trained and tested a CNN to identify various retinal features on 2D OCT images from patients with DME.
Methods :
1,896 2D OCT images through the fovea from 93 patients with DME were randomly sampled from a publicly available dataset of 108,316 2D OCT images. Each image was labeled with the following 5 features: hyperreflective foci, ellipzoid zone disruption, subretinal fluid, vitreoretinal interface abnormalities, and distortion of the foveal contour. An EfficientNet-B3 CNN was trained to identify these features. The training process was repeated 3 times with separate random seeds, resulting in an ensemble of 3 CNNs. Inference was subsequently performed on an independent test set of 250 B-scans from 167 patients with DME by taking the average prediction scores across the 3 CNNs. The area under the receiver operating characteristic curve (AUC) was calculated for each feature. The 95% confidence interval (95% CI) was calculated using the bootstrap method.
Results :
The distribution of the 5 features in the training set was: hyperreflective foci (37.1%), ellipsoid zone disruption (23.8%), subretinal fluid (5.2%), vitreoretinal inferface abnormalities (62.3%), and distortion of the foveal contour (50.9%). Among the 5 features, performance (AUC, 95% CI) was highest for subretinal fluid, 0.947 (0.914, 0.974); hyperreflective foci, 0.918 (0.877, 0.954); and distortion of the foveal contour 0.887, (0.802, 0.957). AUCs for ellipzoid zone disruption and vitreoretinal inferface abnormalities were 0.757 (0.693, 0.819) and 0.738 (0.643, 0.824), respectively. Figures 1 and 2 depict class activation maps illustrating which areas of the image contributed most to the predicted features.
Conclusions :
CNNs were effective in identifying ancillary DME features in 2D OCT images, particularly hyperreflective foci, subretinal fluid, and distortion of the foveal contour. This can allow for more nuanced assessment of DME severity and prognosis in settings where retina subspecialists are not readily available.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.