Enhancing the accuracy of Alzheimer's disease diagnosis through the application of deep learning algorithms for early detection

Abstract

Recently, there has been interest in applying deep learning algorithms to identify Alzheimer’s disease (AD) in its early stages using MRI. During our research, we implemented and benchmarked three deep learning architectures: a 3D convolutional neural network (CNN), a hybrid CNN and long short-term memory (LSTM) network, and a 3D residual neural network (ResNet). A total of 6,400 MRI images covering four stages of AD were used to train and evaluate the models. In the present study, our most optimized and best-performing model, the 3D ResNet, was able to attain an average accuracy of 53.64% in classifying all AD stages. Nonetheless, the model performed well in distinguishing mild to moderate dementia cases, while non-demented and very mild dementia identification was not achieved with an early-stage predictive model. The research was hindered by several essential factors, including class imbalance issues and the model's limited capacity to address different stages of AD. We conclude that deep learning may enhance the accuracy of diagnosing Alzheimer’s disease; however, significant improvements are still needed before it can be applied in clinical practice. It is recommended that multimodal, longitudinal designs and other biomarkers be utilized in future studies to improve diagnostics.