Large-scale neuroimaging data is becoming increasingly available, providing a rich data source with which to study neurological conditions. In this thesis, I demonstrate the utility of large-scale neuroimaging as it applies to Alzheimer’s disease (AD) and normal aging, using univariate parametric mapping, regional analysis, and advanced machine learning. Specifically, this thesis covers: 1) validation and extension of prior studies using large-scale datasets; 2) AD diagnosis and normal aging evaluation empowered by large-scale datasets and advanced deep learning algorithms; 3) enhancement of cerebral blood volume (CBV) fMRI utility with retrospective CBV-fMRI technique.
First, I demonstrated the utility of large-scale datasets for validating and extending prior studies using univariate analytics. I presented a study localizing AD-vulnerable regions more reliably and with better anatomical resolution using data from more than 350 subjects. Following a similar approach, I investigated the structural characteristics of healthy APOE ε4 homozygous subjects screened from a large-scale community-based study. To study the neuroimaging signatures of normal aging, we performed a large-scale joint CBV-fMRI and structural MRI study covering age 20-70s, and a structural MRI study of normal aging covering the full age-span, with the elder group screened from a large-scale clinic-based study ensuring no evidence of AD using both longitudinal follow-up and cerebrospinal fluid (CSF) biomarkers evidences.
Second, I performed deep learning neuroimaging studies for AD diagnosis and normal aging evaluation, and investigated the regionality associated with each task. I developed an AD diagnosis method using a 3D convolutional neural network model trained and evaluated on ~4,600 structural MRI scans and further investigated a series of novel regionality analyses. I further extensively studied the utility of the structural MRI summary measure derived from the deep learning model in prodromal AD detection. This study constitutes a general analytic framework, which was followed to evaluate normal aging by performing deep learning-based age estimation in cognitively normal population using more than 6,000 scans. The deep learning neuroimaging models classified AD and estimated age with high accuracy, and also revealed regional patterns conforming to neuropathophysiology. The deep learning derived MRI measure demonstrated potential clinical utility, outperforming other AD pathology measures and biomarkers. In addition, I explored the utility of deep learning on positron emission tomography (PET) data for AD diagnosis and regionality analyses, further demonstrating the broad utility and generalizability of the method.
Finally, I introduced a technique enabling CBV generation retrospectively from clinical contrast-enhanced scans. The derivation of meaningful functional measures from such clinical scans is only possible through calibration to a reference, which was built from the largest collection of research CBV-fMRI scans from our lab. This method was validated in an epilepsy study and demonstrated the potential to enhance the utility of CBV-fMRI by enriching the CBV-fMRI dataset. This technique is also applicable to AD and normal aging studies, and potentially enables deep learning based analytic approaches applied on CBV-fMRI with similar pipelines used in structural MRI.
Collectively, this thesis demonstrates how mechanistic and diagnostic information on brain disorders can be extracted from large-scale neuroimaging data, using both classical statistical methods and advanced machine learning.
Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/d8-sshc-zj10 |
Date | January 2019 |
Creators | Feng, Xinyang |
Source Sets | Columbia University |
Language | English |
Detected Language | English |
Type | Theses |
Page generated in 0.002 seconds