Premium
Early detection of mild cognitive impairment with convolutional neural network based on brain structural connectivity
Author(s) -
Chen Qianyun,
Yan Taiyu,
Abrigo Jill,
Chu Winnie C.W.
Publication year - 2021
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.052845
Subject(s) - diffusion mri , white matter , artificial intelligence , neuroimaging , convolutional neural network , pattern recognition (psychology) , spatial normalization , computer science , tractography , fractional anisotropy , alzheimer's disease neuroimaging initiative , nuclear medicine , magnetic resonance imaging , cognitive impairment , medicine , cognition , psychology , neuroscience , voxel , radiology
Background Structural connectivity derived from diffusion tensor imaging (DTI) reflects white matter network changes during Alzheimer’s Disease (AD) progression, which can potentially serve as a biomarker for prediction. In this study we explored the ability of structural connectivity to predict mild cognitive impairment (MCI) by implementing a convolutional neural network (CNN). Method 613 DTI scans (194 controls, 291 MCI and 128 AD) from 206 patients (113 male, 93 female; age range: 55‐91 years) from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) Phase GO and 2 were included in this study. The clinical diagnosis at last ADNI follow‐up visit served as reference for prediction (average: 44.4 ± 26.7 months). DTI scans were acquired with 5 b 0 and 41 diffusion weighted volumes with b=1000 s/mm 2 at 2.7×2.7×2.7mm 3 resolution. Preprocessing included eddy current correction and co‐registration to normalized T1‐weighted scans. Whole brain white matter fiber was tracked by deterministic tractography using DSI Studio under Q‐space diffeomorphic reconstruction scheme, which calculates the quantitative anisotropy mapping prior to normalization to standard space. An Automated Anatomical Labeling atlas 2 with 120 region‐of‐interests (ROI) was used to parcellate the brain, where each ROI represented a node of the network. The weight of the edge was calculated by number of tracts connecting two ROIs and normalized by medial length. Therefore a 120×120 structural connectivity matrix was generated for each scan. 80% of data was used for training and 10% of data for validation. The parameters that achieved highest accuracy on validation set was applied to the remaining 10% (hold‐out test set) for final evaluation. A seven‐layer CNN model was constructed for differentiating three groups. The performance was evaluated by accuracy and sensitivity. Result On an average of 3.7 years before the diagnosis, the prediction accuracy was 86.9% in validation set and 85.7% in hold‐out test set. The sensitivity to predict controls and MCI were 95.0% and 93.3%, respectively. Conclusion Using CNN, structural connectivity allowed good ability to predict MCI from cohorts at an early stage before diagnosis.