BG-GAN: Generative AI Enable Representing Brain Structure-Function Connections for Alzheimer's Disease

TL;DR

This work addresses the challenge of representing brain structure-function connections in Alzheimer's disease by learning bidirectional mappings between structural and functional brain networks using a novel BG-GAN framework. The approach introduces InnerGCN to fuse multimodal graph information and a Balancer to stabilize GAN training, enabling simultaneous learning of SC2FC and FC2SC mappings with complementary features. Empirical evaluation on ADNI demonstrates that generated connectivity improves AD identification and reveals that structure and function relate but do not exhibit a strict one-to-one correspondence, highlighting a complex brain coordination mechanism. The findings suggest that brain structure provides a functional basis while iteratively shaping functional interactions, with implications for multimodal biomarkers and disease progression analysis.

Abstract

The relationship between brain structure and function is critical for revealing the pathogenesis of brain disorders, including Alzheimer's disease (AD). However, mapping brain structure to function connections is a very challenging task. In this work, a bidirectional graph generative adversarial network (BG-GAN) is proposed to represent brain structure-function connections. Specifically, by designing a module incorporating inner graph convolution network (InnerGCN), the generators of BG-GAN can employ features of direct and indirect brain regions to learn the mapping function between the structural domain and the functional domain. Besides, a new module named Balancer is designed to counterpoise the optimization between generators and discriminators. By introducing the Balancer into BG-GAN, both the structural generator and functional generator can not only alleviate the issue of mode collapse but also learn complementarity of structural and functional features. Experimental results using the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset show that both generated structure and function connections can improve the identification accuracy of AD. The experimental findings suggest that the relationship between brain structure and function is not a complete one-to-one correspondence. They also suggest that brain structure is the basis of brain function, and the strong structural connections are majorly accompanied by strong functional connections.

Figures (10)

• Figure 1: The proposed BGGAN involves two domain: structural domain and functional domain. BGGAN contains two extractors, two generators, two discriminators and two Balancer modules. The training steps are as following: (1) Firstly, extract the brain structural features based on sMRI and DTI data with structural extractor, and extract the functional features based on fMRI data with functional feature's extractor. (2) Secondly, input the structural and functional features to structural and functional classifiers. The classification results are used to reflect the ability of the feature extractor and guide the learning of the extractors. (3) Thirdly, the structural and functional generators use the structural and functional feature matrices from the first step to perform bidirectional mapping between structural and functional domain. (4) Fourthly, the module Balancer merges the structural and functional connection matrices to reduce the gap between the source data and the target data. (5) Fifthly, structural and functional discriminators evaluate the differences between the generated data and the source data. The above steps describe the process from structure to function, and then reverse the generated function back to the structure. In fact, there is also the process of generating from function to structure, and then from structure to function, which is the same as before.
• Figure 2: Classification performance based on different graph convolution methods. GCN, GAE and GAT set the brain structural information as the edges of graph and the brain functional information as the node features of graph. The classification result based on our method reaches best.
• Figure 3: (a) shows the MSE loss between the target and the source domain connections. The orange curve is the loss outputted by the generator without module Balancer, while the green curve is the loss with module Balancer. (b) shows the generated structural connections from the functional generator. The top five pictures are the results from the generator without module Balancer and the bottom five are the results from the generator with module Balancer.
• Figure 4: Statistical analysis for structural and functional brain connections under each category. The figure (a) shows the number trend of structural connections and the figure (b) shows the number trend of functional connections.
• Figure 5: Similarities and differences between brain structural connections and brain functional connections in each category. The brighter the plot, the higher the probability of presence of brain connections. The yellow areas indicate a higher likelihood of brain connections, while the rest of the purple colors indicate that brain connections are less likely to be present. Top five figures show brain connections where structural connections usually exist but functional connections do not. The five figures in the middle show the brain connections that have a high probability of function but rarely structure. And the bottom five show brain connections that both exist multiple times in structure and function.