| CPC A61B 5/055 (2013.01) [G01R 33/4806 (2013.01); G06N 3/042 (2023.01); G06T 7/0012 (2013.01); G06T 2207/10088 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30016 (2013.01)] | 8 Claims |
|
1. A brain atlas individualization method based on magnetic resonance and a twin graph neural network, comprising the following steps:
(1) obtaining magnetic resonance data of subjects, the data being time series data based on apexes of cerebral cortices;
(2) extracting a feature from the magnetic resonance data of the subjects by utilizing functional connectivity RSFC based on a region-of-interest, at the same time, performing Fisher transformation on the extracted feature to normalize the feature, performing exponential transformation on the extracted feature to sparsify the feature, and using the sparsified data as a feature input of the twin graph neural network;
(3) obtaining connection information of surfaces of the cerebral cortices of the subjects according to the magnetic resonance data of the subjects, and calculating an adjacent matrix of each subject as a graph input of the twin graph neural network according to a connection relationship of the apexes of the cerebral cortices described by the connection information;
(4) extracting a central point sampling mask of a group atlas by utilizing a Floyd-Warshall algorithm as a weighting coefficient of a loss function of the twin graph neural network; in a process of selecting the group atlas to calculate the sampling mask, for functional magnetic resonance imaging fMRI cortex surface apexes of any given subject, calculating a shortest path distance SPD from a given apex in the adjacent matrix to other fMRI cortex surface apexes according to the Floyd-Warshall algorithm, and taking a maximum SPD as a centrifugal degree of the given apex; and for each region-of-interest, sorting the centrifugal degrees from small to large, selecting first 20% points with the minimum centrifugal degree as points with high confidence coefficients, and extracting the central point sampling mask;
(5) adding a difference between the subjects to the network loss function, and training the network by adopting a manner of semi-supervised learning with the central point sampling mask and the group atlas as labels, wherein when the loss function of the twin graph neural network is calculated, data of two subjects and a label of whether the subjects belong to a same subject are provided in each input, and the corresponding loss function comprises cross entropy of the group atlas with individualized brain atlases of the two subjects and a contrast loss function between the individualized brain atlases; a weight ratio is 1:1:λ, wherein λ is a hyper-parameter, represented as a weight of the contrast loss function; and a complete loss function of a set of subjects is:
 wherein three items in L represent cross entropy of a probability yi,k that a brain region label of an ith apex is k with brain region predicted values pi,k,m, pi,k,n of a subject m and a subject n, and the contrast loss function ContrastLoss respectively; ContrastLoss represents a measurement of a similarity of pi,k,m, pi,k,n; and wi is a sampling mask of the ith apex, label=1 represents that a set of input data belongs to the same subject, label=−1 represents that a set of input data belongs to different subjects, and threshold is a threshold value, representing that penalty is only performed when a similarity of the different subjects exceeds the threshold value; and
(6) giving any magnetic resonance data, and inputting the data to the trained twin graph neural network after a feature extracting process same as step (2), wherein a specific process is: ChebNet graph convolutional layers are added into a network framework, in the first layer of ChebNet, a number of filters of is 64, an order is 6, regularization is adopted as a training parameter, in the second layer of ChebNet, a number of filters is determined by a selected reference group atlas, an order is 6, and then a Softmax layer is connected; in order to avoid over-fitting, a dropout layer is connected behind each ChebNet layer; in order to construct the twin graph neural network, the network of this part is set as parameter sharing, at the same time, a binary set is input, and the network is trained by adopting a manner of semi-supervised learning with the sampling mask and the group atlas as labels; and the network is trained by using an Adam optimizer, during training, a model with a minimum corresponding loss function value is reserved, and one-hot encoding matrices of individualized partitions output by the twin graph neural network are mapped at a position of a maximum value as one-dimensional vectors along an encoding dimension to obtain individualized brain atlases corresponding to the magnetic resonance data of the subjects.
|