Deep Learning and Ensemble Method for Optic Disc and Cup Segmentation.

Kim J, Tran L, Peto T, Chew EY, HN

IEEE CIBCB 2022, August 15-17, 2022. https://doi.org/10.1109/CIBCB55180.2022.9863022.

---

Abstract:

Glaucoma is a chronic retinal disease that gradually damages the optic nerve. It is a leading cause of irreversible loss of vision. In ophthalmic fundus images, the cup to optic disc ratio measured around the optic nerve is a key measure used to screen for glaucomatous damages. Unfortunately, there is high subjectivity among ophthalmologists in estimating this ratio due to challenges in making reliable disc and cup measurements. To minimize this, we propose an automatic method using deep learning and ensemble method to segment the optic disc and cup. The proposed method comprises two steps. The region of interest (ROI), where optic disc is centered, is detected from a fundus image, following which the optic disc and cup are segmented from the ROI. Mask R-CNN algorithm is used to estimate the ROI, and two ensemble models based on three fully convolutional networks are used for the segmentation of optic disc and cup in parallel. The proposed method is trained and evaluated using the RIGA dataset that contains 750 fundus images and the REFUGE database containing 400 fundus images. The results demonstrate that the proposed method has a better performance compared with the current state-of-the-art algorithms. Our best segmentation results for optic disc shows 0.9303 Jaccard Index (JI) and 0.9635 Dice Coefficient (DC). The best segmentation results for cup shows 0.8096 JI and 0.8915 DC. The average cup to optic disc ratio error shows 0.0429.

Kim J, Tran L, Peto T, Chew EY, HN. Deep Learning and Ensemble Method for Optic Disc and Cup Segmentation. 
IEEE CIBCB 2022, August 15-17, 2022. https://doi.org/10.1109/CIBCB55180.2022.9863022.

URL: https://doi.org/10.1109/CIBCB55180.2022.9863022