Predicting Brain Function from Structure with MRI and Supervised Machine Learning

Abstract

The main goal of this thesis is to develop machine learning based tools that can be used to investigate the relationship between brain structure, with a focus on white matter tracts, and function. In particular, this work introduces novel methodologies that enhance our ability to map and analyze brain connectivity, as well as predict functional outcomes based on structural data. The structure of our brain presents both commonalities and distinctions across its two hemispheres. Understanding these similarities and differences is essential for examining brain connections. As an initial contribution, this thesis introduces CoBundleMAP, an algorithmic pipeline that employs manifold learning to provide a parameterization of the fibers traversing both hemispheres of the brain. It extends our earlier method, BundleMAP, by ensuring correct correspondences between the two hemispheres, and providing two-dimensional parameterizations for sheet-like bundles. This permits a richer understanding of the brain's connectivity and asymmetries. When tested on brain data from 145 individuals, CoBundleMAP proved superior in detecting nuanced group differences compared to its predecessors, and in supporting classification of individuals. The remaining two contributions sought to predict brain functions based on brain connectivity. The first approach harnessed supervised machine learning to discern the connection between connectional fingerprints as captured through Diffusion Spectrum Imaging (DSI), and brain function, in particular language production, as captured by Transcranial Magnetic Stimulation. A key technical contribution of this work is to propose a novel Bag of Features representation of streamlines inferred from DSI, which permits their processing with a linear SVM. Our findings suggest that the link between a brain region’s connections with other parts of the brain and the role of that region in language production can be used to predict, at a level significantly above random chance, the outcomes of TMS experiments. We also provide visualizations that permit a certain level of interpretation. The final contribution extends this overall framework for predicting function from structure in multiple ways. First, instead of TMS, it considers functional MRI (fMRI), a much more widely used method for mapping brain function, for which a substantial amount of data is available from the Human Connectome Project. Second, in addition to the Bag of Features approach, Gaussian Mixture Models and Fisher vectors are employed for a more detailed representation. While attempts to predict fMRI activations from diffusion MRI had been made previously, our proposed representations express a brain region’s connectional fingerprint without requiring a predefined parcellation of the brain, which saves a non-trivial computational step, and permits a more fine-grained analysis. Lastly, we again incorporate an explainable AI component by visualizing the connection clusters that are most relevant for a prediction. The visuals are derived from a median curve of all the connections in a cluster, providing a clearer and simplified view. Through these methods and visual tools, we support comprehending the features that result from brain scans and how they correlate to its functions.