Enhancing retinal disease diagnosis through ablation studies: a robust deep-learning CNN model with improved image quality and generalizability

Abstract

Retinal disease is a global health crisis that impacts millions of individuals annually. Early illness diagnosis is an important field of medical research since eye problems can greatly affect patients' eyes and overall well-being. A plethora of data for model training and improved model designs have led to the rapid development of deep learning techniques for computer vision applications. This research will describe a powerful convolutional neural network (CNN) model that uses ablation studies to categorize eye disease illnesses. The disease is colored, and the overall quality is increased by reducing noise, improving the image, and removing artifacts. The ODIR 5k dataset of retinal images is freely available to the public. Due to improvement techniques, the quantity of images depicting eye diseases has increased. At the outset, the enhanced dataset suggests a basic CNN model. The robust CNN model, which I propose, was derived from an ablation investigation. To train the model, a library of publicly available images depicting eye illnesses is utilized. In terms of accuracy (92%), area under the curve (AUC) (99%), and KAPPA (91%), the recommended robust model outperformed the competition. The model is robust and performs admirably when applied to fresh data. The model also achieves good recall and precision when used to identify binary and multi-class eye disorders.