Bayesian Convolutional Neural Network for Detection of Cataract Ocular Disease

Abstract

Ocular disease diagnosis using fundus images is one of the most challenging tasks in the medical field but is necessary for early screening and treatment. This manual process is extremely time consuming, complex and error prone. Currently there is an increased demand for the utilization of deep learning techniques for the automated detection of ocular diseases, especially for use on biomedical images. However, these conventional techniques such as Deep Neural Networks and Convolutional Neural Networks (CNNs) present some challenges, such as its tendency to overfit on smaller datasets and its inability to measure the uncertainty of its predictions. This is crucial in the medical field to determine the reliability of the predictions made by the automated systems. Thus, in this paper a Bayesian Convolutional Neural Network (BCNN) is implemented for Cataract disease detection to provide the reliability (uncertainty) estimates sought after in the medical field. The BCNN is benchmarked against the implementation of a standard Convolutional Neural Network. The BCNN model was compiled using the negative log-likelihood loss function and an Adam optimizer with a learning rate of 0.001 trained over 100 epochs. The CNN was compiled similarly except for the loss function which was the categorical cross entropy loss. The test results indicate that the BCNN model achieved 93.16% accuracy, while the standard CNN achieved an accuracy of 95%. Both models achieved comparable accuracy results to existing studies that utilized CNN architectures to predict ocular diseases. Although the CNN gave a slightly better accuracy, it cannot account for the uncertainty measurements of its predictions, would be more useful for ophthalmologists.

https://doi.org/10.59200/ICONIC.2024.008