Stroke is one of the leading causes of both death and disability in the world. Consequently, the processes underlying motor recovery are a hot research topic. Electroencephalography (EEG) and diffusion weighted magnetic resonance imaging (dMRI) are two modalities that can be used to find functional and structural predictors for this motor recovery, respectively. Specifically, EEG measures the sources of activity (dipoles) in the brain while dMRI provides estimates of the properties of white matter (WM) tracts such as the fiber orientation. The estimated fiber orientations can be used to reconstruct WM connections in the brain by performing fiber tractography.

In this thesis, we aim to introduce a framework for model selection and probabilistic tractography with parsimonious model selection. Practically, we use a range of multi-tensor models to cope with regions with multiple fiber populations. Furthermore, our probabilistic tractography uses the Cram\'er-Rao lower bound to capture the uncertainty in the fiber orientations. We mitigate the effect of overfitting by using a model selection method that incorporates the ICOMP-TKLD criterion to determine the most appropriate tensor model in each voxel. Ultimately, this framework can be applied to data from stroke patients and combined with functional regions obtained from EEG.

We assessed the performance of the model selection method by investigating the influence of b-value and noise on the ability to detect crossing fibers in the fibercup phantom and human data. In the phantom, our model selection reconstructed all the crossings for the b-value combination of 1500 and \SI{2000}{\s\per\mm\squared} and at a signal-to-noise-ratio (SNR) comparable to clinical acquisitions. Moreover, our model selection method was able to identify the crossing of the corpus callosum and corticospinal tract in the human data.

A range of step sizes and curvature thresholds was used to investigate the sensitivity of our tractography to its input parameters. In general, a smaller step size and lower curvature thresholds resulted in more deterministic behavior, while a larger step sizes and higher curvature thresholds led to more probabilistic behavior and deeper propagation into the gray matter in human data.

We compared the performance of our framework and the open source diffusion MRI toolkit Camino on the fibercup phantom and healthy control data. In this comparison, our framework performed better in curved bundles and reconstructed more lateral projections of the corpus callosum.

Lastly, we explored the subdivision of the brain into modules for stroke patients and healthy controls, by combining our framework with sources obtained from EEG. Fewer modules were found in the patient group, which might be attributed to a change in structural connections after stroke.

Altogether, we have shown that our framework was able to select the appropriate diffusion models in crossing fiber regions and track across these crossings both in a phantom and human data. Furthermore, we demonstrated that it is feasible to combine our framework with source locations obtained from EEG.