To determine the performance of a deep learning model of neural networks in discriminating glaucomatous optic nerves from healthy discs based on digital fundus images alone or in addition to other relevant biological data.

Our dataset consisted of a total of 2,390 fundus images including 605 normal discs and 1,785 glaucomatous discs. The dataset was randomly divided into two parts: 75% of images were used for training and 25% for testing the algorithm. In the deep learning method features are selected and extracted in an automatic way through convolutional layers which extracts features step by step by creating feature maps and convolutional weights in each layer. At the final layer, the feature vector is created and classified using two fully connected layers. The convolutional architecture was first pre-trained using a public dataset with size 422*422 images. When the network achieved the initial weights properly and became ready, study images described above were fed into the network. In the last layer, the feature vector of biological parameter of the relevant fundus image was attached to the feature vector extracted through the convolutional layers. The network was run with learning rates of 0.1, 0.01 and 0.001. The algorithm was developed in Python 3.6 – Anaconda environment and was executed using a system configuration consisting an Intel CPU 4790, 3.6 GHz, 16 GB RAM with GPU.

The highest performance was obtained using six layers of CNNs at a learning rate of 0.001 resulting in 98.22% accuracy, 95.86% sensitivity and 97.06% specificity. Other learning rates missed some efficient and valuable data resulting in less accuracy.

The artificial intelligence model employed in this study was able to discriminate glaucomatous discs from healthy ones with high accuracy. The performance of the neural network was further improved implementing a novel architecture of layers representing biological parameters. The deep learning method using convoluted neural network seems a promising modality in machine learning for detecting glaucoma. Implications of the study include glaucoma screening based on fundus images with or without the inclusion of other biologic data both in a clinical setting or using telemedicine.