CN-121998990-A - Brain network analysis method based on space-time double-dimension multi-scale

CN121998990ACN 121998990 ACN121998990 ACN 121998990ACN-121998990-A

Abstract

The invention relates to the technical field of medical image diagnosis, and particularly discloses a brain network analysis method based on space-time double-dimension multi-scale, which comprises the steps of obtaining functional magnetic resonance images and extracting time sequences of brain regions; A functional brain map is built on fine, middle and coarse multi-scale based on pearson correlation and functional homogeneity weighting, multi-scale space features are extracted through a mixed structure of a U-Net encoder and a multi-scale map isomorphic network, short-term instantaneous fluctuation and long-term trend dependence are respectively captured by adopting an internal and external double-branch selective state space model framework aiming at brain area BOLD signals, time features are obtained through multi-scale context extraction and pathological time sequence attention mechanism reinforcement disease related time sequence fragments, and a classification result is obtained by inputting the time-space features into a classifier after residual gating fusion. The invention breaks through the limitation of single scale and single time sequence modeling, realizes the refinement and functional fusion of the brain space-time characteristics, and improves the accuracy and the interpretability of auxiliary diagnosis of brain diseases.

Inventors

QU AIXI
Wu Yandai
Qiao Lishan
ZHANG LIMEI
LI WEIKAI
LIU YANAN
XU PEIMING

Assignees

山东建筑大学
泰山体育产业集团有限公司

Dates

Publication Date: 20260508
Application Date: 20260410

Claims (9)

1. A brain network analysis method based on space-time double-dimension multi-scale is characterized by comprising the following steps: S1, acquiring a brain function magnetic resonance image of a tested; S2, constructing a brain function network fusing functional homogeneity constraints on multiple scales of thin, medium and thick based on a time sequence of each brain region in the brain function magnetic resonance image; S3, extracting multi-scale space features from a brain function network fused with functional homogeneity constraints by utilizing a multi-scale map isomorphic network; S4, processing the time sequence of each brain region by using a double-branch time sequence module based on a state space model, extracting multi-scale time sequence characteristics, and weighting the multi-scale time sequence characteristics by using a pathological time sequence attention mechanism to obtain weighted time sequence characteristics; S5, fusing the multi-scale space features and the weighted time sequence features to obtain combined space-time features; s6, inputting the combined space-time characteristics into a classifier, and outputting a brain function state classification result.
2. The method for analyzing the brain network based on the space-time two-dimensional multi-scale according to claim 1, wherein in S2, constructing the brain function network fusing functional homogeneity constraints comprises the following steps: calculating the Pearson correlation coefficient between the time sequences of each brain region to obtain a correlation matrix under each scale ; Based on a preset brain function network map, acquiring weight vectors of each brain region belonging to different function networks through a U-Net encoder; The cosine similarity of weight vectors between each brain region pair is calculated to obtain a functional homogeneity matrix under each scale ; For correlation matrix With functional homogeneity matrix Weighting fusion is carried out to obtain an adjacent matrix of the brain function network with fusion function homogeneity constraint : ; Wherein: Is a weighting coefficient.
3. The method for analyzing the brain network based on the space-time two dimensions and multiple dimensions according to claim 2, wherein in S3, extracting the multi-scale spatial features by utilizing the isomorphic network of the multi-scale map comprises the following steps: inputting the functional brain network adjacency matrix with the fine scale into a U-Net encoder to obtain feature mapping corresponding to three spatial scales of fine, medium and coarse; the feature mapping of each scale is respectively input into an independent graph isomorphic network, and topological features of corresponding scales are extracted; and merging topological features output by the isomorphic networks of the graphs, inputting the merged features into a U-Net decoder for up-sampling, and outputting multi-scale spatial features.
4. The brain network analysis method based on space-time two-dimensional multi-scale according to claim 1, wherein in S4, the two-branch time sequence module based on the state space model comprises a high-resolution branch and a low-resolution branch, the high-resolution branch and the low-resolution branch adopt selective state space models to conduct sequence modeling, the high-resolution branch is used for directly inputting the time sequence of each brain region into a first state space model to extract high-resolution time sequence characteristics representing short-term fluctuation, and the low-resolution branch is used for inputting a second state space model after downsampling the time sequence of each brain region to extract low-resolution time sequence characteristics representing long-term trend.
5. The method for brain network analysis based on space-time two-dimensional multi-scale according to claim 4, wherein in S4, weighting the multi-scale temporal features by using a pathological temporal attention mechanism comprises: fusing the high-resolution time sequence features and the low-resolution time sequence features to obtain primary fused time sequence features; Extracting multi-scale context features from the primary fusion time sequence features by using a plurality of one-dimensional convolution layers with different convolution kernel sizes, and fusing to obtain the multi-scale time sequence features; and performing pathological time sequence attention weighting on the multi-scale time sequence features, namely obtaining time sequence importance scores through a linear layer and a Sigmoid activation function, and multiplying the time sequence importance scores and the multi-scale time sequence features element by element to obtain weighted time sequence features.
6. The brain network analysis method based on space-time two-dimensional multi-scale according to claim 1, further comprising introducing functional homogeneity constraint loss in a model training process, wherein the functional homogeneity constraint loss is used for constraining brain region feature vectors learned by a multi-scale map isomorphic network to maximize cosine similarity between the brain region feature vectors and prototype feature vectors of the functional network.
7. The method for analyzing the brain network based on the space-time two dimensions and the multiple dimensions according to claim 1, wherein in the step S5, a residual gating fusion mechanism is adopted for the step of fusing the multiple dimensions and the weighted time sequence features.
8. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the method according to any one of claims 1-7 when the program is executed by the processor.
9. A computer-readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

Description

Brain network analysis method based on space-time double-dimension multi-scale Technical Field The invention relates to the technical field of medical image processing and computer-aided diagnosis, in particular to a brain network analysis method based on space-time double dimensions and multiple dimensions. Background Autism spectrum disorder is a common neurological disorder, and is mainly manifested by social communication disorder, notch behavior and the like. At present, clinical diagnosis mainly depends on behavior observation, and has the problems of strong subjectivity and easy misdiagnosis. Resting state functional magnetic resonance imaging can reflect the connection strength between different functional areas of the brain noninvasively by measuring blood oxygen level dependent signals, and has become an important tool for researching brain diseases and performing auxiliary diagnosis. The traditional brain network analysis method generally builds a functional connection network based on the Pearson correlation coefficient, but the method can only capture the statistical synchronism of brain intervals, is easy to be interfered by noise, and lacks consideration of the inherent functional organization structure of the brain, so that the built brain network has fuzzy functional semantics. In the feature extraction stage, the existing method has two major limitations that in the space dimension, a single-scale graph neural network is adopted, local fine connection and global topological dependence are difficult to model at the same time, in the time dimension, the cyclic neural network is difficult to process long-range dependence, the computation complexity of a transducer model is high, and a single state space model can model the short-term fluctuation and the long-term trend of a brain signal in a mixed mode, different physiological meanings of the brain signal cannot be distinguished, and the capturing of a disease-specific time sequence mode is inaccurate. Therefore, an analysis method capable of constructing a brain network with a higher biological significance and simultaneously finely modeling the brain space-time characteristics is urgently needed to improve the accuracy and the interpretability of auxiliary diagnosis. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a brain network analysis method based on space-time double-dimension multi-scale, which optimizes brain network construction by introducing functional homogeneity constraint, and designs a multi-scale space feature extraction module and a double-branch time sequence feature extraction module which work cooperatively to realize deep fusion and efficient analysis of brain space-time features. The invention is realized by the following technical scheme: The brain network analysis method based on space-time double-dimension multi-scale comprises the following steps: S1, acquiring a brain function magnetic resonance image of a tested; S2, constructing a brain function network fusing functional homogeneity constraints on multiple scales of thin, medium and thick based on a time sequence of each brain region in the brain function magnetic resonance image; S3, extracting multi-scale space features from a brain function network fused with functional homogeneity constraints by utilizing a multi-scale map isomorphic network; S4, processing the time sequence of each brain region by using a double-branch time sequence module based on a state space model, extracting multi-scale time sequence characteristics, and weighting the multi-scale time sequence characteristics by using a pathological time sequence attention mechanism to obtain weighted time sequence characteristics; S5, fusing the multi-scale space features and the weighted time sequence features to obtain combined space-time features; s6, inputting the combined space-time characteristics into a classifier, and outputting a brain function state classification result. Further, in S2, constructing a brain function network that fuses functional homogeneity constraints, comprising: calculating the Pearson correlation coefficient between the time sequences of each brain region to obtain a correlation matrix under each scale ; Based on a preset brain function network map, acquiring weight vectors of each brain region belonging to different function networks through a U-Net encoder; The cosine similarity of weight vectors between each brain region pair is calculated to obtain a functional homogeneity matrix under each scale ; For correlation matrixWith functional homogeneity matrixWeighting fusion is carried out to obtain an adjacent matrix of the brain function network with fusion function homogeneity constraint: ; Wherein: Is a weighting coefficient. Further, in S3, extracting the multi-scale spatial features using the multi-scale map isomorphic network includes: inputting the functional brain network adjacency matrix with the fine scale into a U-Net enc