A Stable Federated Deep Learning Model for The Effective Detection of Ocular Pathology

Abstract

Identifying eye problems from fundus images presents a significant medical challenge. Every pathology exhibits different periods of severity, which can be inferred by identifying the physical traits of the lesions and establishing their presence. Furthermore, many lesions associated with different diseases have common characteristics. Numerous methods for identifying eye diseases from fundus images have been published. Deep learning (DL) based techniques perform better in detection than other methods because they can customize the network to the desired detection output. In this work, DL-based methods are used to identify eye disorders. First, the image is resized, cropped, reflected, and noise-free using the adaptive Wiener filter. SMOTE is employed to deal with data imbalances andaugment data. Lastly, Federated Deep Learning (FDL) is proposed as a sickness identification approach. Four pre-trained models—the convolutional neural network (CNN), VGG16, VGG19, and ResNetV2—were applied using FDL. The learning outcomes of the federated framework are compared with the centrally taught DL models.