Authors: Y. Vasudha, B. Nikhitha
Abstract: Success in treating and preserving eyesight depends on early diagnosis of cataracts, which are a leading cause of reduced vision and blindness worldwide. It takes skilled ophthalmologists to manually evaluate retinal photographs, which may become laborious when screening a large number of individuals. Thanks to recent breakthroughs in Deep Learning, intelligent computer-aided diagnosis systems have been created that can accurately diagnose retinal disorders from medical pictures. This study presents a Vision Transformer (ViT-B16) architecture-based automated system for cataract classification that can distinguish between normal and cataract retinal images. According to the proposed method, a dataset of 1,120 retinal images, distributed equally between normal and cataract categories, is used. Preprocessing steps include scaling and cropping the images, segmenting the dataset, and adding data via zooming and rotating methods before training the model is applied to the retinal photographs. The generalisability of models and picture quality are both improved by this. Different input image sizes (224 × 224 × 3), learning rates (0.009 and 0.001), optimisers (Adam, RMSprop, and SGD), and epochs (25 and 50) are examined in six separate experimental conditions to evaluate the classification model. We maintain a batch size of 32. You may measure performance using many measures such as Accuracy, Precision, Recall, F1-score, and Validation Loss. The sixth training scenario yielded the best experimental results in terms of accuracy (92.86%), precision (100%), recall (85.72%), F1-score (92.31%), and validation loss (0.2504). Users may enter retinal pictures and get automated cataract categorisation results via an interactive interface. The web application is based on Django and contains the trained Vision Transformer model. The proposed framework provides a trustworthy and efficient computer-aided diagnostic method that maintains the original classification scheme, which may assist in early cataract detection in clinical practice.
DOI: http://doi.org/10.5281/zenodo.21375443
