Brain function network classification method based on variational auto-encoder

Abstract

The invention discloses a brain function network classification method based on a variational autoencoder. The method comprises the following steps: The method comprises the following steps of: acquiring T1 weighted MRI and rs-fMRI of a plurality of normal people and patients with brain cognitive impairment; carrying out pretreatment; carrying out double regression analysis by taking the preprocessed rs-fMRI as a regression dependent variable and the brain function network as a regression independent variable to obtain an individual level brain function network; constructing a deep variationalautoencoder (VAE) model, taking the obtained individual level brain function network diagram as the input and output of the VAE, and taking the encoder part as a feature extraction module for obtaining the implicit code of the individual function network; constructing a multi-layer sensor network to classify the codes obtained by the VAE in the step 4; and deducing samples in the test set by using the trained classifiers for different brain function networks, and fusing deduction results of the classifiers to obtain a final classification result.acquiring T1 weighted magnetic resonance imagesT1 Weighted MRI and resting state functional magnetic resonance images rs-of a plurality of normal persons and patients with brain cognitive impairment; fMRI; carrying out pretreatment; pretreated rs- Performing double regression analysis by taking fMRI as a regression dependent variable and taking the brain function network as a regression independent variable to obtain an individual level brainfunction network; constructing a depth variation auto-encoder (VAE) model, taking the obtained individual level brain function network diagram as input and output of the VAE, and taking the encoder part as a feature extraction module for obtaining hidden codes of the individual function network; constructing a multi-layer perceptron network to classify the codes obtained by the VAE in the step 4;inferring samples in the test set by utilizing a plurality of trained classifiers for different brain function networks, and fusing inference results of the plurality of classifiers to obtain a finalclassification result; according According to the invention, the classification accuracy is improved.

Description

technical field [0001] The invention relates to the technical field of medical image analysis, in particular to a method for processing and classifying functional magnetic resonance images in the field of brain science, in particular to a brain function network classification method based on a variational autoencoder. Background technique [0002] The human brain is one of the most complex systems in nature, and various neurophysiological activities of the brain are the result of interactions between different brain regions. The realization of a complex cognitive task often requires the cooperation of multiple brain regions. The study of the interdependence between different brain regions will help to deepen human understanding of the way the brain works. Modeling brain function from the perspective of network has been proved to be an effective research method, and the analysis of brain function network has become one of the important research directions in the field of brai...

AI Technical Summary

Problems solved by technology

[0006] In order to overcome the problem that the input feature vector contains insufficient information in the traditional brain function network classification method, ignores the topological structure relationship between different brain regions in the brain function network, and the traditional machine learning algorithm has limited data modeling capabilities, the present invention The purpose is to propose a brain function network classification method based on variational autoencoders. Group independent component analysis (Group ICA) can effectively distinguish signals with physiological significance from noise signals, and can separate brain networks with co-activity patterns. Network, and then use the double regression method to extract the individual level brain function network; Group ICA as a data-driven method does not need to define the template network in advance, while retaining the topological structure information of different brain regions; variational autoencoder as a deep On the one hand, the neural network generation model can capture the activation degree of different brain regions in the brain network and the non-local relationship between different brain regions, and on the other hand, it can learn the distribution of hidden variables to obtain a set of hidden variables with good distribution properties.

Images