Functional Connectivity Network Learning And Its Applications In Brain Disease Identification

Human brain is the most complex biological system in nature,whose neurons are connected with each other through synapses,forming a highly complex network of brain connectivities.More and more evidences show that the analysis of brain connectivity network not only becomes an important way to understand the working mechanism of brain,but also provides a new perspective to explore the biomarkers of the neurological or psychiatric diseases.Up to now,researchers have proposed numbers of brain network estimation methods,such as full correlation(e.g.,Pearson correlation coefficient)、partial correlation、regularized partial correlation(e.g.,sparse representation),etc.Despite their effectiveness in some practical applications,these methods still include some advantages.Firstly,due to the interference of physiological factors(e.g.,the breathing and heartbeat of the subjects)and the inevitable slight head movement during the scanning process,the obtained magnetic resonance image data may contains the structural noise.Thus,it can be determined that the estimated brain connectivity network-based the above methods may not reliable.Secondly,an ideal brain connectivity network should be sparse and low-rank,but most of the networks based on the above methods contain redundant connections.Even if the network is sparse,the artificial hyperparameters(e.g.,the threshold or regularization term)also need to be select.Based on this,this dissertation focuses on the above two problems,and obtains the corresponding research results:1)An adaptive brain connectivity network learning model based on the transformation space is proposed for improving the quality of data by transforming the data into a lowdimensional space.In addition,the model using a signle framework makes it possible to complete data denoising and brain connectivity network estimation simultaneously.In order to verify the performance of our method,we conduct the experiments to identify the mild cognitive impairment(or autism)from normal controls,and the experimental results show that the proposed method performs better than the baseline method.2)The brain connectivity network learning method based on the non-hyperparameter model is proposed,which not only effectively circumvents the problem of hyperparameter selection,but also obtains sparse network automatically.To verify the reliability of our method,we also conduct two group of experiments(i.e.,the mild cognitive impairment(or autism)from normal controls),and the experimental results show that the proposed method based on seven performance indices obtains the comparable classification performance,even without any hyperparameters.