Brain Functional Parcellation Methods Based On Swarm Intelli- Gence Algorithms

The brain is the smallest,most complex and intelligent system that has been discovered by far.Functional research of the brain is an important and cutting-edge research content in brain science.As a mainstream neuroimaging technology for aquiring brain function data,functional magnetic resonance imaging(fMRI)provides a powerful data support for the study of brain function because of non-invasiveness,high spatiotemporal resolution and easy operation.Brain functional parcellation studies brain functional organization by dividing cerebral cortex,which is a fundamental research method of brain function.At present,most of the brain function parcellation methods based on fMRI data are the application of existing classic clustering methods in it and can not deal with high dimensional and low signal-to-noise ratio of fMRI data,showing some shortcomings of weak search capability,noise sensitivity and poor functional consistency and regional continuity of parcellation structures.In turn,swarm intelligence algorithms have strong global search ability and some robustness,and show better performance than classic clustering algorithms in clustering.With the respect to the above drawbacks in the study of brain functional parcellation method,this thesis first reviews brain functional parcellation based on fMRI data.And then,on the basis of swarm intelligence algorithms,this thesis conducts innovative research on static brain functional parcellation methods and dynamic brain functional parcellation ways.The specific research work is as follows:1.In the overview of brain functional parcellation oriented to fMRI data,fMRI data acquisition,the basic concept,the classification and the basic process of brain function parcellation oriented to fMRI data are introduced.And then,brain functional parcellation methods based on fMRI data are elaborated from a computational model or mechanism perspective,where a classification system for static brain functional pacellation methods is given.Next,some common functional consistency measures and evaluation indexes are combed.At last,the drawbacks in brain functional parcellation are deeply analyzed.2.Aiming at the problem that fMRI data is low in signal-to-noise ratio and the expected maximum algorithm is easy to get into local optimal when searching Gaussian mixture model(GMM),an insula functional parcellation method based on GMM searched by immune clonal selection(ICS)algorithm is put forward.The proposed method first maps GMMs onto antibodies and then the optimizing search of GMMs is achieved by simulating the immune mechanisms composed of clonal antibodies,clonal mutation and clonal selection.Among them,the clonal mutation adopts a hybrid mutation strategy that flexibly takes different ways of mutation according to the number of iteration stagnation,improving the search ability of ICS.Meanwhile,dynamic neighborhood information with anti-noise capability are employed in the searching process,effectively reducing the adverse effects of noise in fMRI data.Finally,insula functional parcellation is obtained by employing the obtained optimal Gaussian mixture model according to maximum a posteriori criterion.The experimental results suggest that the new method can not only search a better GMM,but also obtain a parcellation structure with strong functional consistency and regional continuity in comparison with some other brain functional parcellation methods.3.As for high dimensionality and low signal-to-noise ratio of fMRI data,a brain functional parcellation method based on artificial bee colony(ABC)algorithm.This new method first projects preprocessed fMRI data into a low-dimension space to reduce its dimension,and then some parameters and population are initialized.Among them,an individual in population is set to a vector,representing a cluster solution.The optimizing search process of cluster solutions is completed by four search mechanisms to simulate the bee colony foraging behavior: self-adaptive crossover search,employed bee search,onlooker bee search and scout bee search.Among them,the proposed selfadaptive crossover mechanism simulates the foraging behavior that a queen bee organizes and coordinates bee colony and its specific operation is in the following: each individual and the current best one are crossed in line with their fitness,overcoming the lack of information exchange among individuals and enhancing the diversity of population.The stepwise search for onlooker bees take advantage of both intermediate and final calculation results in one search,enhancing the search width of onlooker bees and the diversity of candidate individuals in the whole search process.Finally,the labels of data points are obtained according to the principle of the minimum within-cluster distance and mapped onto voxels for brain functional parcellation.The experimental results on a simulated fMRI data set suggest that the novel method can yield the closest to the division of a real result.Compared to some other brain functional parcellation methods,the results on real fMRI data show that the new method not only has stronger search capability,but also can obtain parcellation structures with better functional consistency and regional continuity.Furthermore,the rationality of parcellation results also is verified by connectivity fingerprints of corresponding subregions.4.With respect to the parameter configuration of sliding windows and the low efficientcy of clustering algorithms in the stuty of dynamic brain functional parcellation methods,a dynamic brain functional parcellation method based on sliding window and artificial bee colony algorithm,which composes of sliding-window length determination phase,functional state identification phase and functional parcellation phase.In the first phase,the newly proposed functional connectivity similarity minimum criterion is employed to determine the length of a sliding window.Then,time series are windowed,and functional connectivity matrices measuring function between voxels in a parceled region and them in its ipsilateral hemisphere is calculated for each windowed time series.In the second phase,some functional states are identified by clustering these matrices with an improved ABC.In the improved ABC,employed bee search based on hybrid strategy takes the original search or the optimal solution corresponding bits fill according to a predefined probability threshold,improving the diversity of employed bee search and the "pull" effect of the optimal solution.According to a constraint radius computed with population distribution,the dynamic radius-constrainted scout bee search forces scout bees to search outside the area that is centered on the abandoned food source and bound by this radius,increasing the global search ability of scout bees.In the third phase,functional connectivity between voxels in each function state is computed by connecting time series belonging to the same state in chronological order,and functional parcellation results for all functional states are achieved by the improved ABC.Finally,in comparison with two dynamic brain functional parcellation methods,the experimental results on two real fMRI data sets both suggest that the novel method has obvious advantages in terms of search capability and parcellation structures and reveal the dynamic of posterior cingulate cortex.The above research work in this thesis,on the one hand enriches the methodology study of brain functional parcellation based on fMRI data,further deepen people's understanding of the functional organization of the brain and provides the useful methodological assistance for the prevention and diagnosis of brain diseases.On the other hand,the work also broaden the application field of swarm intelligence algorithms and promotes their research and development.Therefore,the researches in this thesis both promot the development of brain functional parcellation research and have the potential practical value.