Classification of Alzheimer’s Diseases’ MRI Brain Images Leveraging 3D Convolutional Neural Networks

Alzheimer’s Disease (AD), the most preponderant type among neurodegenerative diseases, targets the elderly population with symptoms of memory loss, language impairment, and cognitive function degradation. Currently, the disease progression is irreversible. However, early AD detection at the pre-onset stage can decelerate disease development and improve patients’ quality of life. From a radiological perspective, the progression of AD can be predicted on structural magnetic resonance imaging (MRI) images with parenchyma loss in the hippocampal area and ventricular enlargement. Nevertheless, interpreting AD based on MRI remains a significant challenge due to the insufficiency of medical expertise and human effort. Thus, this work aims to build a computer-aided diagnosis system (CAD) that harnesses three-dimensional deep learning architectures. Particularly, the project aims at solving the classification problem on brain images of the normal control group (NC), early mild cognitive impairment (EMCI), and AD patients. The proposed method is implemented on Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset with four stages, including (1) MRI image preprocessing with bias correction, skull stripping, and brain co-registration. (2) Model training of 3D versions of ResNet, EfficientNet, and ShuffleNet. (3) Horizontally stacking the abovementioned models for better classification results. Furthermore, gradient-weighted class activation mapping (Grad-CAM) was applied to visualize and rationalize the model’s decision. The result shows that the stacking model achieves 96% accuracy on three-way classification and 97% on pairwise NC versus AD classification. In addition, the Grad-CAM’s activation distribution focuses on the brain ventricle and hippocampus region.