Cerebellar Partitioning Based On Spectral Clustering And Its Application In Feature Extraction Of Parkinson’s Brain Network

Functional magnetic resonance imaging(f MRI)technique is a powerful tool for depicting brain function and structure,exploring individual differences in brain function,understanding the working mechanisms of the brain,and studying variations in disease features.Proper brain partitioning can assist researchers in summarizing the functional organization of the brain,reducing data dimensions,and enhance the applicability of various models to brain imaging data.In recent years,with the deepening understanding of cerebellum research,it has been found that its functions involve not only motor control but also activation responses in many advanced cognitive functions.However,currently popular methods of cerebellar partitioning have limitations,which to some extent,hinder further detailed research on the cerebellum and cause difficulties in studying cerebellarspecific changes in diseases such as Parkinson’s disease.Therefore,this thesis aims to develop a more stable and refined method of cerebellar partitioning,and apply it to the construction of cerebellar network models and the study of abnormal characteristics of cerebellar networks in Parkinson’s disease.In the first section of this thesis,a new cerebellar partition was constructed and evaluated using spectral clustering and clustering integration procedures,by utilizing the function-connection matrix as a similarity measure optimized by a monotonic nonlinear transformation.Initially,the new partition was demonstrated to have high inter-individual reproducibility through NMI and Dice coefficients.Additionally,when compared to four publicly available templates,the new partition showed superior signal consistency and greater spatial similarity with the typical cerebellar structure.Furthermore,the granularity of the partition can be adjusted by modifying the number of clusters to accommodate different levels of precision requirements for subsequent Parkinson’s cerebellum network analysis research.The novel approach presented in this study provides a foundation for cerebellar network research and has the potential to be extended to other,similar studies in the field.The cerebellum is an important diseased area of Parkinson’s disease,but the limited division of the cerebellar region restricts further research.The construction of the above new division compensates for this deficiency.The second study in this thesis applied the new division to extract features of functional connections in the Parkinson cerebellar network.Before applying it,the stability of the partitioning method was first demonstrated on different datasets.Based on the above series of partitions and four publicly available cerebellar templates,features of the cerebellar connectivity network were extracted,and a Parkinson’s disease classification model was constructed using a logistic regression model with L2 regularization.The results showed that,compared with the publicly available cerebellar templates,brain functional connection features extracted using the new partitioning greatly improved the availability of the Parkinson’s classification model.Furthermore,cerebellar areas associated with motor execution and other functions exhibited higher feature importance in the Parkinson classification model.This provides an important direction for the selection of candidate brain regions for subsequent multimodal Parkinson’s classification models.The edge-centric functional network provides a new understanding of brain function from the perspective of overlapping functional networks,but the premise of such models also requires appropriate brain regions to define nodes.The first outcome of this thesis improves the feasibility of applying this model to the cerebellum.In the third part of the thesis,the above-mentioned new partition was used to divide nodes in the cerebellum and define the edge time series and edge correlation matrix.By reducing and clustering the edge correlation matrix,a cerebellar overlapping community model was obtained,and the normalized mutual information entropy was used to quantify the degree of functional overlap among cerebellar regions.After applying this procedure to data from Parkinson’s disease patients and healthy control group,it was found that the three overlapping functional communities identified by this method were able to well reproduce the default mode network and two motor representation areas of the cerebellum,and it was discovered that the functional overlap from top to bottom in the cerebellum showed a trend of first increasing and then decreasing.In inter-group difference studies,it was found that the prefrontal and posterior lobes of the cerebellum in Parkinson’s disease patients exhibited higher functional overlap,while the vermis region located in the middle of the cerebellum exhibited the opposite characteristics.These results provide more insights into our understanding of cerebellar network function from the perspective of the edge-centric functional network as well as Parkinson’s cerebellar network variation.