Premium
Detecting biological changes in longitudinal MRI scans of Alzheimer’s disease patients in the hippocampus region with deep learning
Author(s) -
Dong Mengjin,
Xie Long,
Das Sandhitsu R.,
Wolk David A.,
Yushkevich Paul A.
Publication year - 2020
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.046453
Subject(s) - temporal lobe , hippocampus , magnetic resonance imaging , medicine , biomarker , artifact (error) , neuroscience , alzheimer's disease , hippocampal formation , neuroimaging , nuclear medicine , psychology , disease , pathology , radiology , epilepsy , biology , biochemistry
Abstract Background Alzheimer's Disease (AD) is characterized by gradual loss of gray matter, particularly in the medial temporal lobe (MTL) and hippocampus. Longitudinal change in hippocampal volume on MRI is an effective biomarker to monitor disease progression in AD. However, neurodegenerative changes in the hippocampus on longitudinal MRI can be obscured by differences in MRI signal that are unrelated to disease progression, such as different amounts of head motion, noise, or MRI artifact. We hypothesize that a deep learning model can be trained to distinguish true biological changes from nonsystematic factors, leading to a more sensitive biomarker of disease progression. Method A 3D convolutional neural network (CNN) was trained to predict the inter‐scan interval (in days) from pairs of rigidly co‐registered longitudinal T1‐weighted MRI scans. Scans trimmed to the hippocampus region were input in arbitrary time order, and the network had to infer which was the earlier scan and the interscan interval in days. A cohort of 666 ADNI participants was used for CNN training and 154 participants formed a held‐out test set. We tested the hypotheses that (a) CNN can correctly predict the temporal order of scans (indicating sensitivity to biological changes) and (b) that a measure of mismatch between actual interscan interval and interval predicted by a CNN trained on more impaired subjects (late MCI and AD) would differ between less impaired (controls + early MCI) and more impaired groups (indicating sensitivity to AD). The same hypotheses were also tested using a conventional technique for longitudinal change estimation, ALOHA (Das et al., 2012). Result CNN was more accurate than ALOHA in inferring the temporal order of scans (Table 2, Figure 1) and to a lesser extent, inter‐scan interval (Table 3, Figure 2). Mismatch measure was greater for the less impaired than more impaired group for both CNN and ALOHA, with larger difference for CNN, potentially indicating greater sensitivity to disease. Conclusion CNN can accurately detect the temporal order of scans indicating sensitivity to neurodegeneration and ability to factor out non‐systematic changes. The CNN‐derived mismatch measure has the potential to serve as a disease progression biomarker for AD clinical trials.