Segmentation of Optic Nerve images for glaucoma detection, using U-Net Deep Learning Model

Abstract
Purpose : To evaluate de U-Net Deep Learning ((Convolutional Networks for Biomedical Image Segmentation) for automatic segmentation of optic disc and their disc cupping using eye fundus images.

Methods : The training of the Deep Learning model was performed using 209 eye fundus pictures of glaucomatous optic nerves and healthy optic nerves. The Deep learning model was tested using 26 images which had been previously evaluated by a glaucoma specialist, who delineated the border of the optic nerve and the internal border of the rim using Labelbox tool.
Finally, the Jaccard coefficient was used to evaluate cuantitavely the accuracy of the model segmentating the image of the optic nerve and the cupping of the disc to detect glaucoma.

Results : The results of the evaluation show a Jaccard index of 0.82 for the segmentation of the optic disc and 0.72 for the segmentation of the disc cupping.

Conclusions : Although the results are aceptable, a study with a larger number of images is needed to determine the use of U-Net in the detection of glaucomatous optic nerves. U-Net model has shown good results with smaller samples compared to other Deep Learning models, which need larger samples.