Abstract
Purpose :
Efforts at automatic identification of the optic disc and its neural rim have been reported, but have not been used to grade glaucoma severity with an accepted metric. In this retrospective observational study, we propose a DDLSNet pipeline consisting of a rim segmentation (RimNet) branch and a disc size classification (DiscNet) branch to automate glaucoma grading with the disc damage likelihood scale (DDLS).
Methods :
To develop RimNet, funduscopic images from the UCLA Stein Glaucoma Division were randomly split into training, validation, and test datasets (80/10/10). Optic rims of the training set were drawn by glaucoma specialists. RimNet uses contrast enhancement, an InceptionV3/LinkNet rim segmentation model, and computer vision to identify the thinnest rim location and calculate the rim-to-disc ratio (RDR). To develop DiscNet, paired funduscopic images and OCT data from UCLA Stein Glaucoma Division were randomly split into training, validation, test datasets (80/10/10). DiscNet uses a VGG19 network to label disc size as small, medium, or large. The DDLS grade is calculated from RDR and disc size. Three glaucoma specialists graded the RimNet test set with DDLS (scale 1-10). The main outcome measure was a weighted kappa agreement with agreement defined as +/- 1 DDLS grade between graders and DDLSnet.
Results :
RimNet was developed on 857 images (mean age=60.3 (±14.2) years, male:female ratio=0.42), and achieved an RDR mean absolute error of 0.06 (±0.04) on test set between RimNet and physician. DiscNet was developed on 8366 images (mean age=67.6 (±14.5) years, male:female ratio=0.57), and achieved 77% classification accuracy on test set. DDLSNet was evaluated on 87 images from the RimNet test set; Table 1 lists the agreements between DDLSNet and graders and between individual graders.
Conclusions :
DDLSNet achieved acceptable performance as shown by fair weighted kappa agreement with grader 2 and 3. This illustrates the feasibility of developing an automated process such as DDLSNet for glaucomatous grading of optic disc images. Further improvements are required, and will be achieved by further training of graders, increasing sample size, and algorithm enhancement.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.