Research On Brain Fingerprinting Recognition Based On Functional Magnetic Resonance Imaging

The brain fingerprinting based on resting-state functional magnetic resonance imaging(rs-f MRI)refers to the existence of unique features in the rs-f MRI signal,which can be used to indicate the uniqueness of an individual.However,the most relevant features about brain fingerprinting still have no uniform definition standard.The release of the Human Connectome Project,the development of machine learning and deep learning have laid a technical foundation for the exploration of brain fingerprints.Brain fingerprinting recognition based on rs-f MRI mostly uses all static functional connectivity for individual recognition.The feature dimension is high,and there may be redundant features that affect the recognition performance.From the perspective of selecting features that have a greater impact on recognition,a brain fingerprint recognition method based on multi-scale group Lasso is proposed;then in order to further extract brain fingerprint features from dynamic functional connectivity that contain more information,this thesis proposes a brain fingerprinting recognition method based on the combination of BiLSTM network and spatialtemporal attention mechanism.In addition,this thesis applied the SENet network,and proposes a brain fingerprinting recognition method based on the multi-scale spatiotemporal SENet network.Most of the existing researches are based on small sample data,and the sample size is only a few hundred,which is easy to cause over-fitting of the model.This thesis uses a data set containing 1003 samples for experimentation.The specific work is as follows:(1)In this thesis,we propose a method based on multi-scale group lasso to select discriminative static functional connectivity for brain fingerprinting recognition.Most of the existed brain fingerprinting recognition methods are based on all static functional connectivity,and it is difficult to determine which functional connectivity contribute more to brain fingerprinting recognition and affect the recognition performance.The use of group lasso can select the most recognizable functional connection,improve the recognition performance.The proposed method is superior to other methods in performance,and the recognition rate can reach 98.21%.(2)In this thesis,we propose a brain fingerprinting recognition method based on the combination of BiLSTM network and spatialtemporal attention mechanism.Compared with static functional connectivity,dynamic functional connectivity contains more abundant information,which is helpful to improve the performance of brain fingerprinting recognition.Therefore,this thesis calculated the dFC of rs-f MRI based on the sliding window method,extracted the spatial and temporal features of the dynamic functional connectivity using the BiLSTM network,and enhanced the feature representation by assigning different weight coefficients to the features combined with the spatialtemporal attention mechanism,and the recognition rate reached 82.20%.(3)In this thesis,a multi-scale spatiotemporal SENet network is proposed to integrate the spatiotemporal characteristics of multi-scale dFC for brain fingerprinting recognition.Multi-scale dFC,with different step size,contains the complementary information,in order to integrate the spatialtemporal features of multi-scale dFC to utilize more information,this article applies the SENet network to highlight the features which have larger effect on the recognition task,and restrain the features have less identify influence.In other words,different weight coefficients are applied to spatiotemporal features of different scales to improve the recognition performance.The proposed method achieves a recognition accuracy of 92.17%,which is better than other methods.The three proposed methods all improve the recognition accuracy of brain fingerprinting,and help to better promote the development of brain fingerprinting recognition research,thereby providing help for predicting diseases.