Implementation of Deep Learning based Automated Diagnosis of Glaucoma Using Digital Retinal Fundus Images

Abstract

Glaucoma is a prevalent chronic condition that can cause irreversible vision loss. The number of individuals suffering
from permanent vision loss as a result of glaucoma is predicted to rise at an alarming rate in the near future. There
is a lot of study being done on computer-aided diagnosis for glaucoma. The optic cup (OC) and optic disc (OD) are
typically segmented in retinal fundus images to distinguish between glaucomatous and non-glaucomatous instances.
However, the OC boundaries are quite non-distinctive; as a result, accurate OC segmentation is extremely difficult,
and OD segmentation performance also needs to be improved. To address this issue, we suggest two networks for
accurate glaucoma screening: CNN and RNN-LSTM. We created a CNN-RNN hybrid that extracts not only the spatial
information in a fundus image but also the temporal features encoded in fundus sequential images. A CNN and a
combined CNN and Long Short-Term Memory RNN were trained using 1810 fundus pictures and 295 fundus videos.
In differentiating glaucoma from healthy eyes, the combined CNN/RNN model achieved an average F-measure of
95.2%. In comparison, the fundamental CNN model only achieved an average F-measure of 78.2%. Both proposed
networks include a separable convolutional link to improve computational efficiency and lower network costs. The
proposed architecture can provide great accuracy even with only a few trainable parameters.