Author ORCID Identifier

https://orcid.org/0000-0001-9167-3718

Date of Award

8-8-2024

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Jingyu liu

Second Advisor

Sergey Plis

Third Advisor

Jessica Turner

Fourth Advisor

Dong Hye Ye

Abstract

With recent technological advances in acquiring multimodal brain imaging data and high-throughput genomics data, brain imaging genomics is emerging as a rapidly growing research field. This field performs integrative studies that analyze genetic variations such as single nucleotide polymorphisms (SNPs), and brain imaging quantitative traits (QTs), coupled with other biomarkers, clinical, and environmental data. The goal is to gain new insights into the phenotypic characteristics and the genetic mechanisms of the brain, as well as their impact on normal and disordered brain function and behavior. Our research proposes a series of algorithms to analyze imaging genomics. Initially, we investigated the ability to predict the trajectory of symptoms in both inattention and hyperactivity domains using features from sMRI images and genomics of an ADHD cohort of 77 subjects. We performed stepwise linear regression coupled with stability selection and permutation tests to identify the top predictive features. In the inattention domain, age, genes OSBPL1A, CTNNB1, PRPSAP2, ACADM, and one GM component in the insula region were associated with symptom change, while in the hyperactivity domain, no features were associated with symptom change. In our second study, we proposed a strategy for training convolutional neural networks (CNN) with limited samples using a self-transfer-training (STT) method, which refines and reuses layers to optimize model performance. Thirdly, we examined the potential of CNN models trained on structural MRI images to classify working memory capacity and understand the brain regions contributing to memory tasks. A CNN model trained on brain age prediction of 39,755 subjects was transferred to a working memory classification task with fewer subjects, leveraging the features learned on brain age prediction. Lastly, our fourth study integrates neuroimaging and genetics via contrastive learning for working memory prediction. We utilized data from the UK Biobank, combining structural MRI and SNP data with advanced machine learning techniques, including contrastive learning and sparse canonical correlation analysis (sCCA), to uncover significant relationships between genetic variants and brain regions. This integrated approach achieved superior classification accuracy, providing new insights into the genetic and neural mechanisms underlying working memory, and demonstrating the potential of multi-modal data integration in cognitive research.

DOI

https://doi.org/10.57709/37362849

File Upload Confirmation

1

Share

COinS