Early detection of Alzheimer's disease using 3D convolutional neural networks

• Published in: Alzheimer's & dementia Vol. 17; no. S4
• Main Authors: Pan, Dan, Zeng, An, Zou, Chao, Rong, Huabin, Song, Xiaowei
• Format: Journal Article
• Language: English
• Published: 01.12.2021
• ISSN: 1552-5260; 1552-5279
• DOI: 10.1002/alz.053169

Background As one of the most common neurodegenerative disorders which progress slowly over time, Alzheimer's Disease (AD) could be effectively managed by delaying the disease process through early detection and intervention at present. Mild Cognitive Impairment (MCI) is the prodromal state of AD. At present, in‐vivo structural magnetic resonance imaging (MRI) has been widely used for computer‐aided diagnosis of neurodegenerative disorders noninvasively owing to its sensitivity to morphological changes caused by brain atrophy like AD. Plus, the rapid progress of deep learning (DL), especially deep convolutional neural networks (CNNs) has improved MRI analysis thanks to its superiority in the generalization capability. Method An ensemble learning (EL) model which combines genetic algorithm (GA) with 3D‐CNN based on region of interest (ROI) was proposed to identify AD/MCI subjects (Figure 1). We firstly used the 3D‐CNN model to train a candidate base classifier for each a ROI (here a brain region). Then, the GA algorithm was employed to search for the best base classifier combination with the optimal generalization ability and based on it, the whole‐brain MRI classifier ensemble was built to detect AD/MCI. Owing to the one‐to‐one correspondence relationship between the base classifiers and the brain regions, we further identified those brain regions with significant classification capabilities. Result In three binary classification tasks, i.e., classification between 1) AD vs. NC (Normal control), 2) MCIc (MCI patients who will convert to AD) vs. NC and 3) MCIc and MCInc (MCI patients who will not convert to AD), the testing results revealed accuracy rate of 0.89±0.03, 0.88±0.03, and 0.71±0.08, with a stratified fivefold cross‐validation method, respectively. In a data‐driven way, the brain regions that greatly contributed to AD and MCI classifications (Figure 2), e.g., medial amygdala, rostral hippocampus, caudal hippocampus, were ascertained. They were linked to emotion, memory, language, and other key brain functions impaired early in the AD staging. Conclusion Compared with the 2D‐CNN models, the proposed classifiers ensemble could make full use of the effective information embedded in MRI to discover more discriminative MRI features/biomarkers. Additionally, the advocated method could also be valuable to detect the potential neuroimaging biomarkers for other brain disorders.