Early Detection of Glaucoma using Transfer Learning from Pre-trained CNN Models

Abstract

Glaucoma is one of the common diseases that might cause visual field loss, and typically affects elderly people. It is caused by fluid imbalance within the eye that leads to increase in intraocular pressure (IOP), and therefore a damage to the optic nerve head (ONH) which is responsible in transmitting visual neurological signals to the brain. Traditional methods for detecting Glaucoma disease either tedious and slow or too expensive. Hence, early detection of Glaucoma is essential to avoid permanent blindness which might be caused by the ONH failure. In this paper, an automated detection method on the basis of pre-trained Convolutional Neural Network (CNN) models is proposed to detect Glaucoma from fundus images. The proposed method not only contributes to early detection of Glaucoma disease, but also helps optometry doctors in making fast decision with inexpensive tools. Pre-trained AlexNet, VGG11, VGG16, VGG19, GoogleNet (Inception V1), ResNET-18, ResNET-50, ResNET-101 and ResNet-152 models were leveraged to develop the proposed Glaucoma detection method. The proposed method was evaluated by Large-scale Attention based Glaucoma (LAG) dataset. Satisfying results of 81.4%, 80%, 82.2%, 80.9%, 82.9%, 86.7%, 85.6%, 86.2%, and 86.9% were observed on LAG dataset using AlexNet, VGG11, VGG16, VGG19, GoogleNet (Inception V1), ResNET-18, ResNET-50, ResNET-101 and ResNet-152 models respectively. Out of these results, the ResNet-152 model found to be the best that achieved a high accuracy with precision 86.9% and recall 86.9%.