Automated Age-Related Macular Degeneration Diagnosis in Retinal Fundus Images via ViT

Age-related Macular Degeneration (AMD) is a visual impairment condition that commonly affects elderly individuals, and it is a top cause of vision loss in individuals over the age of 60. Early detection can help identify a treatment plan to slow the progress of the disease and prevent severe loss of vision, but this requires comprehensive eye examinations and accurate diagnosis of the disease. The diagnosis is subject to human error. This chapter proposes the use of an automated detection system that circumvents these issues. Multiple deep learning techniques are explored in order to identify an optimum algorithm for automated detection. Deep learning algorithms, particularly Convolution Neural Networks (CNNs), are the front-runners for medical image classification and detection problems. While CNN models are accurate, they are also computationally intensive. This chapter explores the possibility of a lightweight deep learning architecture namely Vision Transformer (ViT) and its variants to diagnose AMD. A comparison is done between state-of-the-art CNN models including AlexNet, MobileNet, and XCeption model and the transformer models including ViT, Modified ViT, and Swin Transformer. The comparison is done on the basis of accuracy, average training time (ATT), CPU utilization, and GPU utilization in order to identify models that can be used for mobile applications with good accuracy.