A new technique for cataract eye disease diagnosis in deep learning

Abstract

Automated diagnosis of eye diseases using fundus images is challenging because manual analysis is time-consuming, prone to errors, and complicated. Thus, computer-aided tools for automatically detecting various ocular disorders from fundus images are needed. Deep learning algorithms enable improved image classification, making automated targeted ocular disease detection feasible. This study employed state-of-the-art deep learning image classifiers, such as VGG-19, to categorize the highly imbalanced ODIR-5K (Ocular Disease Intelligent Recognition) dataset of 5000 fundus images across eight disease classes, including cataract, glaucoma, diabetic retinopathy, and age-related macular degeneration. To address this imbalance, the multiclass problem is converted into binary classification tasks with equal samples in each category. The dataset was preprocessed and augmented to generate balanced datasets. The binary classifiers were trained on flat data using the VGG-19 (Visual Geometry Group) model. This approach achieved an accuracy of 95% for distinguishing normal versus cataract cases in only 15 epochs, outperforming the previous methods. Precision and recall were high for both classes – Normal and Cataract, with F1 scores of 0.95-0.96. Balancing the dataset and using deep VGG-19 classifiers significantly improved automated eye disease diagnosis accuracy from fundus images. With further research, this approach could lead to deploying AI (Artificial intelligence)-assisted tools for ophthalmologists to screen patients and support clinical decision-making.