SFC-GAN: A Generative Adversarial Network for Brain Functional and Structural Connectome Translation

TL;DR

The paper addresses the challenge of jointly analyzing brain structural and functional connectomes by proposing a bidirectional translation framework, SFC-GAN, that extends CycleGAN with two generators ($G_{FC}$, $G_{SC}$) and two discriminators ($D_{FC}$, $D_{SC}$) to translate between FC and SC. It introduces a structure-preserving loss $ ext{L}_{SP}$ that combines $ ext{L}_{MSE}$ and $ ext{L}_{PCC}$ to preserve global and local connectome patterns and symmetry, alongside standard $ ext{L}_{GAN}$, $ ext{L}_{cyc}$, and $ ext{L}_{id}$ terms. The approach demonstrates improved translation quality and topology preservation on ADNI and DUMC-MDD datasets, with translated connectomes retaining discriminative information for downstream SVM classification and preserving key graph properties. This work enables robust cross-modality synthesis when one connectome is missing and provides a framework for topology-preserving connectome translation in brain network analysis.

Abstract

Modern brain imaging technologies have enabled the detailed reconstruction of human brain connectomes, capturing structural connectivity (SC) from diffusion MRI and functional connectivity (FC) from functional MRI. Understanding the intricate relationships between SC and FC is vital for gaining deeper insights into the brain's functional and organizational mechanisms. However, obtaining both SC and FC modalities simultaneously remains challenging, hindering comprehensive analyses. Existing deep generative models typically focus on synthesizing a single modality or unidirectional translation between FC and SC, thereby missing the potential benefits of bi-directional translation, especially in scenarios where only one connectome is available. Therefore, we propose Structural-Functional Connectivity GAN (SFC-GAN), a novel framework for bidirectional translation between SC and FC. This approach leverages the CycleGAN architecture, incorporating convolutional layers to effectively capture the spatial structures of brain connectomes. To preserve the topological integrity of these connectomes, we employ a structure-preserving loss that guides the model in capturing both global and local connectome patterns while maintaining symmetry. Our framework demonstrates superior performance in translating between SC and FC, outperforming baseline models in similarity and graph property evaluations compared to ground truth data, each translated modality can be effectively utilized for downstream classification.

Figures (3)

• Figure 1: Overview of SFC-GAN: a Training of SFC-GAN, $G_{SC}$ translates FC, $\bm{x}_{fc}$ to SC $\tilde{\bm{x}}_{sc}$, where $G_{FC}$ translates SC, $\bm{x}_{sc}$ to FC $\tilde{\bm{x}}_{fc}$. $D_{FC}$ and $D_{SC}$ aim to discriminate between $\bm{x}_{fc}$ and $\tilde{\bm{x}}_{fc}$, $x_{sc}$ and $\tilde{\bm{x}}_{sc}$, respectively. b Cycle consistency loss of FC and SC domains. c Network architectures of $G_{FC}$, $G_{SC}$, $D_{FC}$, and $D_{SC}$.
• Figure 2: Matrix reconstruction results on both ADNI and DUMC-MDD datasets using SFC-GAN.
• Figure 3: Reconstruction results on both ADNI and DUMC-MDD datasets with top 5% strongest connectivity in the brain space.