Abstract
Background: Ocular diseases have been a severe problem worldwide, specifically in underdeveloped countries that do not have enough technology or economy to treat them. It would be beneficial to have software with low installation complexity and ease of use, allowing high efficacy in diagnosing eye diseases. This study aims to design and implement an algorithm based on deep learning to classify ocular diseases with high precision. Methods: This work describes digital image processing techniques for the easier handling of eye images; in particular, blur filters were used. The Canny filter was also applied to obtain the edges that allow the difference between the analyzed diseases. Once the images were pre-processed, a convolutional neural network of our own design was applied to perform the classification task. The validation algorithm used in this work was the hold-out algorithm (80-20). The metrics used to evaluate our proposal were the confusion matrix, accuracy, recall precision, and F1-score. Results: The dataset has five classes, namely, normal, cataract, diabetic retinopathy, glaucoma, and other retina diseases. The network architecture consists of 11 layers, including three convolutional layers, three max pooling layers, one batch normalization layer, one flattening layer, two hidden layers, and one output layer. This model resulted in 97% efficiency across all metrics. Conclusions: With the individual analysis of each metric, it can be observed that the proposed algorithm is capable of differentiating, first, images of healthy eyes from diseased ones and, second, adequately classifying eye diseases.