Machine Learning Techniques for Improved Functional Brain Parcellation
Abstract
Brain parcellation studies are fundamental for neuroscience as they serve as a bridge between anatomy and function, helping researchers interpret how functions are distributed across different brain regions. However, two substantial challenges exist in current imaging-based brain parcellation studies: large variations in the functional organization across individuals and the intrinsic spatial dependence which causes nearby brain locations to have a similar function. This thesis presents a series of projects aimed to tackle these challenges from different perspectives by using advanced machine learning techniques.
To handle the challenge of individual variability in building precise individual parcellations, Chapter 3 introduces a novel hierarchical Bayesian brain parcellation framework. This framework learns a brain probabilistic parcellation by integrating across diverse datasets. For single individuals, the framework optimally combines the limited individual data with the group probability map, resulting in improved individual maps. We found that the resultant individual parcellation based on only 10 minutes of imaging scans can achieve an equivalent performance to the ones using 100 minutes of data alone. These improved individual parcellations are essential to accurately capture functional variations across studied populations.
The intrinsic spatial dependence between brain locations poses a significant challenge in both evaluating and generating brain parcellations. To address this, Chapter 2 presents a bias-free method for evaluating different brain parcellations, the distance-controlled boundary coefficient (DCBC). Compared to existing evaluation metrics that bias toward finer and spatial contiguous parcellations due to spatial smoothness, DCBC provides a fair evaluation by controlling the distance of brain location pairs, ensuring a direct comparison of parcellations in different resolutions. To address the intrinsic spatial dependence when learning parcellations, I propose a new model in Chapter 4, the multinomial restricted Boltzmann machine (m-RBM), that can be incorporated into the learning framework in Chapter 3. This model captures spatial structure between brain locations in its architecture. While simulations showed the utility of this type of model in estimating individual parcellations, we could not demonstrate better performance using real imaging data.
Together, this thesis significantly advances the technical toolkit for deriving brain parcellations from functional imaging data. The developments open up new avenues for future research into human brain organization.