Deep Learning for Brain Structural Connectivity Analysis: From Tissue Segmentation to Tractogram Alignment

Magnetic Resonance Imaging (MRI) is a cornerstone in neuroimaging for studying brain anatomy and functions. Anatomical MRI images, such as T1-weighted (T1-w) scans, allow the non-invasive visualization of the brain tissues, enabling the investigation of the brain morphology and facilitating the diagnosis of both acquired (e.g., tumors, stroke lesions, infections) and congenital (e.g., malformations) brain disorders. T1-w images provide a detailed representation of brain anatomy and accurate differentiation between the main brain structures, such as white matter (WM) and gray matter (GM), therefor they are frequently used in combination with advanced sequences such as diffusion MRI (dMRI) for the computation of the structural connectivity of the brain. In particular, from the processing of dMRI data, it is possible to investigate the structures of WM through tractography techniques, obtaining a virtual representation of the WM pathways called tractogram. Since the tractogram is a collection of digital fibers representing the neuronal axons connecting the brain's cortical areas, it is the fundamental element for studying the brain's structural connectivity. A critical step for processing the tractography data is the accurate labeling of the brain tissues, usually performed through brain tissue segmentation of T1-w images. Even though the gold standard is manual segmentation, it is time-consuming and prone to intra/inter-operator variability. Automated model-based methods produce more consistent and reliable results, however, they struggle with accuracy in the case of pathological brains due to reliance on priors based on normal anatomy. Recently, deep learning (DL) has shown the potential of supervised data-driven approaches for brain tissue segmentation by leveraging the information encoded in the signal intensity of T1-w images. As a first contribution of this thesis, we reported empirical evidence that a data-driven approach is effective for brain tissue segmentation in pathological brains. By implementing a DL network trained on a large dataset of only healthy subjects, we demonstrated improvements in segmenting the brain tissues compared to models based on healthy anatomical priors, especially on severely distorted brains. Additionally, we published a benchmark for enabling an open investigation into improving tissue segmentation of distorted brains, providing a training dataset of about one thousand healthy individuals with T1-w MR images and corresponding brain tissue labels, and a test dataset includes several tens of individuals with severe brain distortions. Another crucial aspect of processing tractography data for brain connectivity analysis is the correct alignment of the WM structures across different subjects or their normalization into a common reference space, usually performed as tractography alignment. The best practice is to perform the registration using T1-w images and then apply the resulting transformation to align the tractography, despite T1-w images lacking fiber orientation information. In light of this, various methods have been proposed to leverage the information of the WM from dMRI data, ranging from scalar diffusion maps to more complex models encoding fiber orientation in the voxels. As a second contribution to the thesis, we provide a comprehensive survey of methods for conducting tractogram alignment. Additionally, we include an empirical study with the results of a quantitative comparison among the main methods for which an implementation is available. From our findings, the use of increasingly complex diffusion models does not significantly improve the alignment of tractograms. Conversely, correspondence methods that use the fibers directly to compute the alignment outperform voxel-based methods, albeit with some limitations: not producing a deformation field, operating in an unsupervised manner, and avoiding using anatomical information. Recently, geometric deep learning (GDL) models have shown promising results in handling non-grid data like tractograms, offering new possibilities for WM structure alignment. The third main contribution of this thesis is implementing a GDL model for tractogram alignment through a supervised approach guided by fiber correspondence. The alignment is predicted as the displacement of fiber points, based on a GDL registration framework that combines graph convolutional networks and differentiable loopy belief propagation, incorporating the definition of fiber structure into the encoding of the graph. Our empirical analysis demonstrates the advantages of utilizing the proposed GDL framework over traditional volumetric registration, showcasing high alignment accuracy, low inference time, and good generalization capabilities. Overall, this thesis advances the methodology for processing MRI data for brain structural connectivity, addressing the challenges of tissue segmentation and tractography alignment, proving the potential of DL approaches also in the case of pathological brains.