Resting-State Functional Magnetic Resonance Imaging-Based Human Brain Parcellation And Its Applications To Brain-Related Diseases

Brain atlas divides the human brain into different subregions according to cell types,anatomical organization,and functional homogeneity,which is an important representation of the brain as an advanced intelligent system.The fine-grained and accurate brain atlas is not only a crucial approach to understanding the brain,but also an urgent requirement for brain-related disease mechanism research,disease diagnosis,and clinical treatment,and to guide and promote the development of brain-inspired intelligence technology.Parcellation technology of resting-state functional magnetic resonance imaging(rs-fMRI)can non-invasively reveal the fundamental organizational principles of the human brain,which is a significant research direction for functional brain atlases.Presently,however,brain functional atlases are mainly derived from static and linear connectivity features of the data drawn from healthy populations,neglecting the effects of brain diseases,individual differences,and the brain’s nonlinear dynamic nature.To address these limitations,we conduct the following studies in this dissertation while considering perspectives of disease-specific atlas,individualized parcellation,nonlinear feature extraction,and dynamic network parcellation.(1)The first Parkinson’ s disease(PD)-specific cortical atlas(PD 186 atlas)was constructed based on a boundary mapping approach using rs-fMRI data from 186 patients with PD from the Parkinson’s Progression Markers Initiative(PPMI)database,and the results were validated and explored in three independent cohorts.The results demonstrated that the PD186 atlas better represented the PD population,with higher boundary similarity and parcel homogeneity.For the spatial topography,the PD 186 atlas contained fewer parcels in the sensorimotor network(hand)while more parcels in the visual network.The size of the central parcel of the primary visual cortex was negatively correlated with disease severity.Additionally,compared to healthy controls,PD patients exhibited reduced functional homogeneity,especially in sensorimotor and visual parcels,which decreased with disease severity.These results validate and highlight the necessity of developing disease-specific atlases.The PD 186 atlas reveals topographic and functional alterations in sensorimotor and visual areas of PD patients,offering potential indicators for disease diagnosis and assessment that will facilitate future mechanism research and clinical exploration of PD.(2)An efficient and flexible framework called AGP is proposed for individualized cortical parcellation.This framework uses a prior brain atlas to initialize seeds,and subsequently adapts a region growing algorithm merging the unassigned and neighbor vertex with the highest-correlated parcel iteratively.The algorithm considers both functional organization distribution of prior atlases and excellent functional homogeneity of individual parcels.The proposed framework was applied to 100 uncorrelated subjects from the Human Connectome Project(HCP)database for functional homogeneity comparison and individual identification,and to 186 PD patients from the PPMI database for symptom prediction.The results demonstrated that the AGP framework outperformed other methods in functional homogeneity,and the generated parcellations provided 100%individual identification accuracy.Additionally,the default mode network(DMN)exhibited higher functional homogeneity,intra-subject parcel reproducibility,and individual identification accuracy,while the sensorimotor network did the opposite.The correlation analysis showed that the disease severity was related negatively to the similarity of individual parcellation and the prior atlases of healthy populations.The disease severity can be correctly predicted using machine learning models based on individual topographic features as well.These results demonstrate that DMN is the most representative,stable,and discriminative network in the resting state.The proposed framework not only improves functional homogeneity but also captures individualized as well as disease-related spatial topological features,providing efficient individualized parcellation tools for future research of brain function and disease.(3)A deep embedded connectivity-based parcellation(DECBP)framework is proposed to leverage deep nonlinear features to delineate fine-grained atlases of brain areas.This framework is the first to introduce deep neural networks to perform individual and group striatal parcellation based on rs-fMRI connectivity.The deep neural network consists of a stacked autoencoder and a clustering module,that can simultaneously and unsupervised extract nonlinear features and perform cluster assignments.The results demonstrated that DECBP exhibited higher intra-subject,inter-subject,and inter-group reproducibility compared to three prevalent clustering techniques and their combination with linear dimensionality reduction methods.Subsequently,the generated striatal atlases were applied to PD(23 patients and 27 healthy controls),and the finer-grained atlas further localized the disease damage to the putaminal subregions,especially the posterior/caudal putamina.Weaker coupling between the putaminal subregions and other brain regions may reflect neuroimaging manifestations of the altered cortico-striatothalamo-cortical circuit and negatively correlated with disease severity.These results validate the reliability of the DECBP framework and demonstrate the nonlinear features improved the robustness of the atlas.The clinical application shows that a finer-grained atlas can more accurately localize the damaged area,and the abnormal connectivity of the putaminal subregions is expected to be neuroimaging indicator for PD.(4)Studying dynamic brain network parcellation under disease based on dynamic functional connectivity.Based on a sliding window approach combined with clustering and community detection techniques,125 subjects(62 children with ASD and 63 healthy controls)from the Autism Brain Imaging Data Exchange(ABIDE)database were employed to delineate representative brain network parcellation under dynamic connectivity,and understand the neural mechanisms of the disease based on connectivity temporal variability and parcellation state configuration.The results showed that,for children with ASD,the functional connectivity between posterior cingulate gyrus and pars opercularis of the inferior frontal gyrus presented greater variability,increasing depending on the disease severity.Additionally,the dynamic network delineated an abnormal overly extensive DMN parcellation merging the areas of DMN,sensorimotor,and central executive networks,associated with a whole-brain hyper-connected pattern.Children with ASD exhibited a higher probability of transition and longer mean dwell times in this parcellation.These results demonstrate that dynamic functional connectivity contains more comprehensive information than static connectivity and is a more sensitive indicator of ASD disease.Furthermore,children with ASD are more dominated by the overly extensive DMN parcellation derived from the hyper-connected pattern,providing new insights for understanding brain parcellation and neural mechanisms in ASD.In summary,this dissertation proposes a series of machine learning frameworks to address the shortcomings of current brain atlas and thus,enhance and improve the robustness and homogeneity of brain parcellation and advance the development of the brain atlas toward disease,individualization,and dynamics.The close association of brain parcellation with disease also provides new perspectives to disease research and promotes the exploration of its neurophysiological mechanisms,which is expected to play an important role in prediction,assessment,diagnosis and treatment of diseases.