Twin Graph Neural Network for Individualized Brain Atlas Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional brain atlas individualization methods lack representation of individual differences, particularly in regions with significant changes in brain size, spatial position, and arrangement, leading to difficulties in behavioral prediction and disease diagnosis, and have slow convergence and long reconstruction times.
Innovation Solution
A brain atlas individualization method and system based on magnetic resonance and a twin graph neural network, utilizing a twin graph neural network with Chebyshev polynomial graph convolutional structure, incorporates high-order domain information and introduces a sampling mask to retain consistency and reflect differences between individuals, employing semi-supervised learning with a group atlas and sampling mask as labels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If traditional linear or nonlinear registration methods are used for individualized brain atlas construction, then the method is relatively simple to implement, but the reconstruction time is long and convergence is slow
Solution Approach 1:
The patent replaces traditional mechanical registration methods (linear/nonlinear registration in Euclidean or cortical space) with a deep learning-based twin graph neural network. This substitution enables parallel processing of brain region correspondence identification across multiple subjects simultaneously, dramatically reducing reconstruction time while maintaining implementation feasibility through standardized network architecture.
Solution Approach 2:
The patent performs preliminary construction of a group atlas that encodes inter-subject variability and intra-subject consistency before individualized atlas reconstruction. This pre-computed group atlas serves as a reference framework that guides the twin graph neural network, enabling faster convergence during individualized reconstruction without requiring complex iterative registration procedures.
2Measurement precision
If group atlas based on averaging many subjects is used, then macroscopic tissue structure analysis is improved, but individual differences in brain size, spatial position and arrangement are not represented
Solution Approach 1:
The patent segments the brain atlas construction process into two distinct components: a group atlas that captures macroscopic tissue structure through averaging, and individualized atlases that capture subject-specific variations. The twin graph neural network separately models inter-subject differences and intra-subject consistency, allowing both group-level and individual-level information to be preserved without mutual interference.
Solution Approach 2:
The patent applies different processing strategies to different aspects of brain anatomy: the group atlas provides standardized macroscopic structure representation, while the individualized atlas components specifically model local variations in brain size, spatial position, and regional arrangement. This local quality differentiation ensures that neither group-level nor individual-level information is lost.
3Device complexity
If traditional methods only use part of the priori knowledge of rs-fMRI signals, then the algorithm is simpler, but the reconstruction accuracy of individualized brain atlases is limited
Solution Approach 1:
The patent combines multiple sources of priori knowledge from rs-fMRI signals into a composite modeling framework: functional connectivity matrices capturing cortical correlations, group atlas encoding inter-subject variability, and intra-subject consistency constraints. The twin graph neural network integrates these diverse information sources through a unified loss function, achieving high reconstruction accuracy by leveraging the complementary strengths of each knowledge source.
Data Source
AI summary
The present disclosure discloses a brain atlas individualization method and system based on magnetic resonance and a twin graph neural network. Firstly, a feature is extracted from resting-state functional magnetic resonance imaging (rs-fMRI) by utilizing functional connectivity based on a region-of-interest, and at the same time, Fisher transformation and exponential transformation are performed on the feature; secondly, a corresponding adjacent matrix is extracted from T1-weighted magnetic resonance data in a data set; and then the twin graph neural network is designed for training and testing with the transformed feature and the adjacent matrix as inputs and a group atlas label and a sampling mask as outputs.