Design And Implementation Of FMRI Subspace Clustering System Based On Auto-encoder

Functional magnetic resonance imaging(fMRI)technology is one of the important technologies for studying human brain function.In recent years,the method of assisting the diagnosis of brain diseases through brain function network connection has attracted the attention of a large number of scholars,and has become a hotspot in medical research on neuropsychiatric disorders.At present,the brain function network analysis strategy widely used in the world is to extract features of brain functional connection data,and then use classification methods to identify diseases.Because of the large dimension of brain functional network data,the auto-encoder is undoubtedly a better choice,and the auto-encoder can not only reduce the dimension,but also extract more abstract features from the training data.In most studies that combine the brain functional network and auto-encoders,they often use auto-encoders to extract features and then use supervised learning algorithms to predict the disease,but this requires labels and is still subjective.Clustering in unsupervised methods can achieve the purpose of predicting disease categories.Therefore,this paper studies the brain function network through auto-encoders and clustering methods.The work is summarized as follows:(1)The traditional brain network analysis methods often use auto-encoders to extract features,and then use supervised learning algorithms to predict diseases.This paper proposes an auto-encoder based multi-kernel fuzzy clustering method(AEMKFC).This AE-MKFC method can not only extract the abstract features of the brain network,but also use the multi-dimensional hidden layers' information on auto-encoder to perform the clustering.A sufficient comparison experiment verifies the effectiveness of the improved AE-MKFC algorithm.(2)For the two-step method of feature extraction of brain functional connectivity data by most researchers,and then using classification methods for diseases identification,this paper proposes an end-to-end unsupervised deep multi-kernel autoencoder clustering network(DMACN).It can not only learn potential advanced features,but also classify diseases at the same time.(3)Designed and developed an fMRI subspace clustering identification system based on an auto-encoder.The system mainly includes three functional modules: fMRI image preprocessing,diseases cluster recognition and performance evaluation.It implements fMRI data preprocessing into corresponding brain functional networks,dimension reduction and clustering the brain functional networks,and measure the algorithms to achieve the purpose of identifying brain diseases and to assist doctors in diagnosis.