Towards Federated-Deep Learning-Based Glaucoma Detection from Color Fundus Images

Glaucoma, a leading cause of irreversible vision loss, can be prevented if detected early. Color Fundus Photography (CFP) is the most efficient and cost-effective method for retinal imaging. Despite the widespread use of deep learning in medical image analysis, its adoption for glaucoma diagnosis has been limited due to insufficient data availability. In this study, we propose a network design consisting of eight convolutional layers, followed by maxpooling and fully connected layers, and a 1000-way SoftMax activation, utilizing the Federated Learning technique. To accelerate the training process, we combine non-saturating neurons with a fast GPU version of the convolution technique. Considering the sensitive nature of medical information, we employ Federated Learning, specifically FedAvg, to ensure data protection. To augment the available training data, we implement random sampling techniques. Our approach employs the AlexNet deep neural network model within the Federated Learning framework. We conduct thirty federated learning iterations with ten clients to train the global model. The final accuracy achieved on the training data is 97.93%, with a test data accuracy of 95.48%. These results demonstrate the effectiveness of our model in glaucoma diagnosis using CFP images. The combination of Federated Learning and deep learning techniques presents a promising avenue for improving glaucoma diagnosis while preserving data privacy.