Automatic Glaucoma Diagnosis in Digital Fundus Images Using Deep CNNs

The World Health Organization (WHO) approximates that more than 42 million people are currently blind in the world, 80 per cent of which could have been prevented or cured by early detection. According to a survey, glaucoma is the second most leading cause for blindness after cataract. It is an irreversible eye disease, and once the vision is lost, it cannot be recovered. Thus, it is vital to develop an automatic computerized tool to diagnose the disease. In this paper, a novel and robust deep learning-based convolutional neural networks (CNNs) architecture has been proposed to deal with the problem. The network consists of six convolutional layers, with various activation functions, and pooling layers to get the abstract and detailed information of the input image. The proposed architecture predicts the probability of an image being glaucoma. The model has been experimented with Refugee and Drishti datasets. Our proposed model is able to diagnose the glaucoma disease automatically with an accuracy of 90%, sensitivity of 96%, and specificity of 84%, respectively.

世界卫生组织（WHO）估计，目前全球有超过4200万人失明，其中80%的失明本可以通过早期检测预防或治愈。一项调查显示，青光眼是仅次于白内障的第二大致盲原因。它是一种不可逆的眼病，一旦视力丧失，无法恢复。因此，开发一种自动的计算机化工具来诊断这种疾病至关重要。在本文中，我们提出了一种新颖且稳健的基于深度学习的卷积神经网络（CNN）架构来解决这个问题。该网络由六个卷积层组成，具有各种激活函数，还有池化层以获取输入图像的抽象和详细信息。所提出的架构可预测图像为青光眼的概率。该模型已在Refugee和Drishti数据集上进行了实验。我们提出的模型能够自动诊断青光眼疾病，其准确率分别为90%、敏感度为96%、特异度为84%。