Glaucoma Classification using Light Vision Transformer

Abstract

INTRODUCTION: Nowadays one of the primary causes of permanent blindness is glaucoma. Due to the trade-offs, it makes in terms of portability, size, and cost, fundus imaging is the most widely used glaucoma screening technique.
 OBJECTIVES:To boost accuracy,focusing on less execution time, and less resources consumption, we have proposed a vision transformer-based model with data pre-processing techniques which fix classification problems.
 METHODS: Convolution is a “local” technique used by CNNs that is restricted to a limited area around an image. Self-attention, used by Vision Transformers, is a “global” action since it gathers data from the whole image. This makes it possible for the ViT to successfully collect far-off semantic relevance in an image. Several optimizers, including Adamax, SGD, RMSprop, Adadelta, Adafactor, Nadam, and Adagrad, were studied in this paper. We have trained and tested the Vision Transformer model on the IEEE Fundus image dataset having 1750 Healthy and Glaucoma images. Additionally, the dataset was preprocessed using image resizing, auto-rotation, and auto-adjust contrast by adaptive equalization.
 RESULTS: Results also show that the Nadam Optimizer increased accuracy up to 97% in adaptive equalized preprocessing dataset followed by auto rotate and image resizing operations.
 CONCLUSION: The experimental findings shows that transformer based classification spurred a revolution in computer vision with reduced time in training and classification.