RimNet: a deep neural network for automated identification of the optic disc edge and neural rim

Abstract

Purpose : Accurate measurement of the neural rim based on optic disc imaging is an important aspect of glaucoma severity grading, often performed best by a trained glaucoma specialist. We aim to improve upon existing partially automated tools by building a fully automated system (RimNet) for direct rim segmentation and rim-to-disc ratio (RDR) calculation in glaucomatous eyes with images obtained in a retrospective observational study.

Methods : An database of 857 glaucomatous funduscopic images of 695 eyes were used for training, validation, and testing divided by an 80%/10%/10% random split. All rims were manually delineated by glaucoma specialists. The dataset consists of 591 patients (242 male, 349 female). Mean age was 60.4 years (±14.2). Any images deemed of insufficient quality by a glaucoma specialist were excluded. RimNet consists of a deep learning rim segmentation model and a computer vision RDR measurement tool. A random search algorithm was used to identify the best performing deep learning architecture for rim segmentation, as determined by intersection-over-union with the clinician segmentations. The measurement tool uses the model-generated rim segmentation to identify the thinnest section of the rim and calculate the RDR. The Drishti-GS dataset was used for an external validation of RimNet performance (Sivaswamy 2015).

Results : RimNet achieved a mean average error (MAE) of 0.06 (±0.04) for RDR on the test set. Figure 1 demonstrates training, validation, and test RDR results. An MAE of 0.13 (±0.09) was achieved for rim-to-disc area ratio (RDAR) on the test set. Bland Altman plots are shown in Figure 2. On the Drishti-GS dataset an MAE of 0.06 (±0.04) for RDR and 0.12 (±0.08) for RDAR was observed.

Conclusions : RimNet efficacious rim segmentation and RDR calculations, as demonstrated by the low MAEs and good agreements on both our test and DRISHTI-GS datasets. Such an automated algorithm would be a valuable component in an automated RDR-based glaucoma grading system. Further improvements could be made by improving the deep learning algorithm and expanding the variety of glaucomatous training images.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Violin plots showing the distributions of clinician and RimNet RDR values in the training, validation, and test images.

 

Bland-Altman plots showing the agreements in RDR and RDAR between clinician and RimNet in test images. Red dashed lines indicate 95% confidence limits.