Repository logo
 

Application of Deep Learning to Brain Connectivity Classification in Large MRI Datasets


Type

Thesis

Change log

Authors

Abstract

The use of machine learning for whole-brain classification of magnetic resonance imaging (MRI) data is of clear interest, both for understanding phenotypic differences in brain structure and function and for diagnostic applications. Developments of deep learning models in the past decade have revolutionized photographic image and speech recognition, bringing promise to do the same to other fields of science. However, there are many practical and theoretical challenges in the translation of such methods to the unique context of MRIs of the brain. This thesis presents a theoretical underpinning for whole-brain classification of extremely large datasets of multi-site MRIs, including machine learning model architecture, dataset curation methods, machine learning visualization methods, encoding of MRI data, and feature extraction. To replicate large sample sizes typically applied to deep learning models, a dataset of over 50,000 functional and structural MRIs was amassed from nine different databases, and the undertaken analyses were conducted on three covariates commonly found across these collections: sex, resting state/task, and autism spectrum disorder. I find that deep learning is not only a method that has promise for clinical application in the future, but also a powerful statistical tool for analyzing complex, nonlinear relationships in brain data where conventional statistics may fail. However, results are also dependent on factors such as dataset imbalances, confounding factors such as motion and head size, selected methods of encoding MRI data, variability of machine learning models and selected methods of visualizing the machine learning results. In this thesis, I present the following methodological innovations: (1) a method of balancing datasets as a means of regressing out measurable confounding factors; (2) a means of removing spatial biases from deep learning visualization methods; (3) methods of encoding functional and structural datasets as connectivity matrices; (4) the use of ensemble models and convolutional neural network architectures to improve classification accuracy and consistency; (5) adaptation of deep learning visualization methods to study brain connections utilized in the classification process. Additionally, I discuss interpretations, limitations, and future directions of this research.

Description

Date

2020-06-01

Advisors

Suckling, John

Keywords

Deep Learning, Functional MRI, Brain Connectivity

Qualification

Doctor of Philosophy (PhD)

Awarding Institution

University of Cambridge
Sponsorship
Gates Cambridge Scholarship