Multi-class Prediction of Cognitively Normal / Mild Cognitive Impairment / Alzheimer's Disease Status in Dementia Based on Convolutional Neural Networks with Attention Mechanism.

Alzheimer's disease (AD) is a neurodegenerative disease with insidious onset and progressive development. AD is a health issue that is attracting attention as the world's populations get older. Although there is currently no effective treatment for this disease, early diagnosis is necessary to help people prevent it. Here, we developed an MRI-based deep learning and multi-class AD classification and prediction framework. Based on the traditional Visual Geometry Group 19 (VGG19) architecture, we embedded the Convolutional Block Attention Module Attention layer to build an Alzheimer's directional prediction 3D convolution model referred to as AD_Net in this paper. We used MRI images from the ADNI open data resource to train, classify, and predict three different conditions: AD, mild cognitive impairment (MCI), and cognitively normal (CN). Experimental results showed that the prediction accuracy and system robustness of AD_Net are superior to those of basic VGG19. To improve the accuracy of the prediction, we developed a multilayer perceptron (MLP)-based model to incorporate additional factors into the model, such as age, gender, and Mini-Mental State Examination score. We presented the model performance for groups with and without directional influence factors. Accuracy for the two groups was 51.2% and 89%, respectively. The latter group had a standard deviation as low as 1.8%, reflecting predictive performance that is both good and stable. Our model can be applied to the early diagnosis of Alzheimer's disease and other similar conditions that can be diagnosed by MRI and the patient's underlying factors.