Date of Award
2-19-2024
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Functional MRI (fMRI) has long been employed to tackle questions related to “how the brain works”. Nevertheless, the data collected through fMRI typically contains a substantial amount of variation due to artifacts and noise, with only a small fraction reflecting genuine neural signals, thus highlighting the need for denoising. Moreover, it is crucial that the fMRI denoising methods be highly accurate by minimizing inclusion of artifacts that could be mistaken for neurobiological signals, while minimizing the removal of genuine neural signals to prevent the potential mischaracterization of brain areas or functional networks. Similarly, when studying parcellation of individual brains into functionally distinct subdivisions, it is essential to ensure that the focus remains on genuine individual cortical areal variability, rather than being confounded by variability caused by the effects of noise on surface registration or automated parcellation. To achieve this, the areal classifier method selected should also maximize both individual variability and test-retest reproducibility. This thesis is dedicated to advancing our ability to understand human brain organization by focusing on two key components: 1) an effective cleaning process to obtain high-quality fMRI data, particularly from the Human Connectome Project (HCP), and 2) the development and validation of an improved individual cortical parcellation method to define the brain areas for each participant within the 1071-subject HCP-Young Adult dataset. In this work, two complementary denoising pipelines were developed to clean both global and spatially specific artifacts from functional MRI data. These pipelines employed Independent Component Analysis (ICA)-based techniques and were automated through the utilization of traditional machine learning and deep learning algorithms. The temporal ICA pipeline primarily focused on the removal of global artifacts, such as respiration, and consisted of nine steps. The key innovations included: 1) the development of a group-level sICA dimensionality selection method for MELODIC’s Incremental Group PCA (MIGP), using multiple Wishart Distributions to fit the unstructured noise, and 2) a hierarchical automated classifier for temporal ICA components, employing handcrafted features and spatial latent features learned from self-supervised learning on spatial maps. The temporal ICA pipeline typically requires a large number of total timepoints across the group being denoised; however, we introduced three decomposition modes, facilitating global artifact removal across varying numbers of subjects (the total timepoints) from single subjects to thousands, thereby accommodating a broad spectrum of use cases. Additionally, we developed a spatial ICA reclean pipeline to correct imperfect component predictions generated by the commonly used sICA+FIX method, leveraging knowledge from the large number of partially hand-labeled sICA components from HCP datasets. Lightweight classifiers were trained on millions of sICA components, with majority voting determining the reclassification of signal and artifact. These pipelines exhibit outstanding performance, achieving close-to-unity Precision-Recall Area Under the Curve (PR-AUC) and Receiver operating characteristic AUC (ROC-AUC) on held-out evaluation datasets, and were both integrated into the HCP pipelines for the broader research community to benefit from. As a byproduct of this research, the fMRI datasets from the Human Connectome Project (HCP) were cleaned via the sICA+FIX, sICA reclean, and temporal ICA pipelines. These meticulously curated datasets are scheduled to be released in Spring 2024, including the Young Adult cohort (1071 subjects for 3T, 175 subjects for 7T, aged 22-35), the Lifespan cohorts with Aging dataset (1798 sessions, ages 36-100+), and Development dataset (1695 sessions, ages 5-21). These invaluable datasets will provide the scientific community with high-quality fMRI data to explore the complexities of human brain functional organization across different stages of life. Further, these methods will be used on future HCP datasets including the Adult Aging Brain Connectome project and the Connectomes Related to Human Disease projects. Building upon these cleaned HCP datasets, I developed a refined individual-subject parcellation method that outperforms existing state-of-the-art methods in terms of test-retest reproducibility and detection of individual variability. Notably, this marks the first instance of applying weak supervised learning into the individual parcellation process, with the HCP-Multimodal Parcellation 1.0 serving as the training labels, which is expected to guide and shape the development of future methods for individual parcellation. This new method has enabled a more comprehensive analysis of individual variability, leading to the discovery of two new findings on the organization of area 55b in the left hemisphere. These findings are: 1) an ‘anterior 55b’ pattern, which is not adjacent to M1 as 55b typically is, and 2) a ‘posterior 55b’ pattern, characterized by its adjacency to M1 and non-adjacency to 8Av, in addition to previously described switched and split 55b patterns. In a total 26.1% of areas 55b in at least one hemisphere in the 1071 HCP-Young Adult datasets were found to be atypical, a substantial increase relative to previously reported results.
Language
English (en)
Chair
David Van Essen
Committee Members
Matthew Glasser