Multisite learning in medical image analysis

Michelle Hromatka
T h e U n i v e r s i t y o f U t a h G r a d u a t e S c h o o l STATEMENT OF THESIS APPROVAL The thesis of Michelle Hromatka has been approved by the following supervisory committee members: P. Thomas Fletcher , Chair ABSTRACT Multisite imaging studies have the potential to accelerate scientific discovery by providing increased sample sizes, broader ranges of participant demographics, and publicly available data. However, failing to address the known nuisance variability across sites, such as
more » ... ross sites, such as scanner type or imaging protocol, reduces statistical power of any analysis performed on the multisite data. In this thesis, I present three contributions to the field of medical image analysis that are designed to reduce this known variability. These contributions include a feature reduction technique for pairwise correlation functional-magnetic resonance imaging (fMRI) data used as features in a multisite support vector machine (SVM), a subject-level network estimation technique for structural magnetic resonance imaging (MRI), and a hierarchical atlas estimation approach that accounts for intersite variability, while providing a global atlas as a common coordinate system for images across all sites. All results are presented on the Autism Brain Imaging Data Exchange (ABIDE) data set which contains resting-state fMRI (rs-fMRI) and structural MRI for 1112 subjects, including both autism and control groups. These methods result in state-of-the-art classification accuracy on the ABIDE data set and increased efficiency in reducing overall MRI data variability.
doi:10.26053/0h-tfp4-wtg0 fatcat:ndcwabpdxrh2xge5lepwejzere