MMHGE: detecting mild cognitive impairment based on multi-atlas multi-view hybrid graph convolutional networks and ensemble learning

• Published in: Cluster computing Vol. 24; no. 1; pp. 103 - 113
• Main Authors: Liu, Jin, Zeng, Dejiao, Guo, Rui, Lu, Mingming, Wu, Fang-Xiang, Wang, Jianxin
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.03.2021; Springer Nature B.V
• Subjects: Artificial neural networks; Autism; Brain diseases; Classification; Cognitive ability; Computer Communication Networks; Computer Science; Construction; Datasets; Deep learning; Ensemble learning; Graph representations; Graphs; Impairment; Machine learning; Magnetic resonance imaging; Medical imaging; Morphology; Neural networks; Neuroimaging; Older people; Operating Systems; Processor Architectures; Quality of life; Graph convolutional networks; MCI detection; Ensemble learning; Multi-atlas multi-view feature representation
• ISSN: 1386-7857; 1573-7543
• DOI: 10.1007/s10586-020-03199-8

Currently, it is still a great challenge in clinical practice to accurately detect the early state of Alzheimer’s disease (AD), i.e., mild cognitive impairment (MCI) including early MCI (EMCI) and late MCI (LMCI). To address this challenge, we propose a new MCI detection framework based on multi-atlas multi-view hybrid graph convolutional networks and ensemble learning. We first construct nine different graphs based on three brain atlases and three morphological measurements using both imaging and non-imaging data of each subject. Then, in order to integrate the information of different graphs and obtain more discriminative feature representations for detecting MCI, we propose a hybrid graph convolutional network method. Finally, a new ensemble learning method is proposed to perform MCI detection tasks. An evaluation of our proposed framework has been conducted with 369 subjects with cognitively normal (CN), 779 subjects with MCI including 310 subjects with EMCI and 469 subjects with LMCI, and 301 subjects with AD on three classification tasks. Experimental results show that our proposed framework can get an accuracy of 90.8% and an AUC of 0.932 for MCI/CN classification, an accuracy of 88.6% and an AUC of 0.908 for MCI/AD classification, and an accuracy of 83.5% and an AUC of 0.851 for EMCI/LMCI classification, respectively. Compared with some state-of-the-art methods about MCI detection, our proposed framework can get better performance. Overall, our proposed framework is effective and promising for MCI detection in clinical practice.