Data-driven methods for characterizing individual differences in brain MRI

Abstract

Understanding the structure and function of the human brain is an outstanding problem that is critical to the development of efficient treatments for neurological diseases like Alzheimer’s and Parkinson’s. While most studies make group level inferences, researchers have established that structure and function show variability across individuals. Motivated by these, recent studies have focused on defining compact characterizations of individual brains, called brain fingerprints. So far, these studies have mostly focused on single modalities, with functional connectivity based fingerprints gaining considerable research interest. However, there are certain aspects which have not been addressed. First, the potential of fingerprints based on structural connectivity has not been fully explored. This is in part due to the challenges arising from fiber tracking data, including lack of gold standard and bundle variability. Second, due to the challenges of combining multiple modalities in a single framework, defining a multi-modal brain fingerprint remains to this day an elusive task. Yet, since each modality captures unique properties of the brain, combining multiple modalities could provide a richer, more discriminative information. This thesis addresses these challenges through three distinct contributions. The first contribution consists of efficient approaches, based on kernel dictionary learning and sparsity priors, for segmenting white matter fibers and characterizing their inter-subject variability. The general principle of the proposed approaches is to learn a compact dictionary of training streamlines capable of describing the whole dataset, and to encode bundles as a sparse combination of multiple dictionary prototypes. These approaches allow streamlines to be assigned to more than one bundle, making them more suitable for scenarios where streamlines are not clearly separated, bundles overlap, or when there is important inter-individual variability. Additionally, they do not require an explicit representation of the streamlines and can extend to any streamline representation or distance/similarity measure. Experiments on a labeled set and data from HCP highlight the ability of our approaches to group streamlines into plausible bundles, and illustrate the benefits of employing sparsity priors. The second contribution is a novel brain fingerprint, called Fiberprint, which is the first to capture white matter fiber geometry in individual subjects. This fingerprint leverages the sparse dictionary learning approaches of the first contribution to map streamlines of subjects to a common space representing prominent bundles. Compact fingerprints are generated by applying a pooling function for each bundle, encoding unique properties of streamlines such as their density along bundles. In a large-scale analysis using data from 861 HCP subjects, the proposed Fiberprint is shown capable of identifying exemplars from the same individual or genetically related subjects, with only a small number of streamlines. Lastly, the third contribution of this thesis is a first data-driven framework to generate brain fingerprints from multi-modal data. The key idea is to represent each image as a bag of local features, and use these multi-modal features to map subjects in a low-dimension subspace called manifold. Experiments using the T1/T2-weighted MRI, diffusion MRI, and resting state fMRI data of 945 HCP subjects demonstrate the benefit of combining multiple modalities, with multi-modal fingerprints more discriminative than those generated from individual modalities. Results also highlight the link between fingerprint similarity and genetic proximity, with monozygotic twins having more similar fingerprints than dizygotic or non-twin siblings. The work described in this thesis can be of benefit to various neuroscience studies. The segmentation approaches presented in the thesis provide a flexible and efficient way to analyze 3D curves like tractography streamlines, and is suitable for large-scale analyses of structural connectivity. The proposed Fiberprint, which is the first brain fingerprint characterizing white matter fiber geometry, offers a powerful technique to explore individual differences in terms of white matter connectivity and its relationship to genetics. By including along-tract information on microstructure, it could also be used to define novel biomarkers for detecting and tracking the progression of neurological diseases like Parkinson’s. Finally, the multi-modal brain fingerprint stemming from this research complements ongoing efforts to analyze individual brains characteristics by allowing to compare and contrast the contribution of different imaging modalities. It can thus lead to new insights on the variability of both brain structure and function, which could help the development of personalized treatment strategies.

Date	11 Jun 2018
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Christian Desrosiers (Supervisor)