[Paper Review] DecGAN: Decoupling Generative Adversarial Network detecting abnormal neural circuits for Alzheimer's disease

Junren Pan, Baiying Lei+4 more|arXiv (Cornell University)|Oct 12, 2021

Cell Image Analysis Techniques52 references4 citations

TL;DR

This paper proposes DecGAN, a novel decoupling generative adversarial network that detects abnormal neural circuits in Alzheimer’s disease by decomposing multimodal brain networks (fMRI and DTI) into sparse subgraphs representing AD-critical circuits and a negligible supplement graph. By integrating a graph-based decoupling module, adversarial training, and a sparse capacity loss grounded in spatial-spectral hypergraph similarity, DecGAN improves detection accuracy and robustness, achieving state-of-the-art performance in classifying AD stages with reconstruction-based brain networks.

ABSTRACT

One of the main reasons for Alzheimer's disease (AD) is the disorder of some neural circuits. Existing methods for AD prediction have achieved great success, however, detecting abnormal neural circuits from the perspective of brain networks is still a big challenge. In this work, a novel decoupling generative adversarial network (DecGAN) is proposed to detect abnormal neural circuits for AD. Concretely, a decoupling module is designed to decompose a brain network into two parts: one part is composed of a few sparse graphs which represent the neural circuits largely determining the development of AD; the other part is a supplement graph, whose influence on AD can be ignored. Furthermore, the adversarial strategy is utilized to guide the decoupling module to extract the feature more related to AD. Meanwhile, by encoding the detected neural circuits to hypergraph data, an analytic module associated with the hyperedge neurons algorithm is designed to identify the neural circuits. More importantly, a novel sparse capacity loss based on the spatial-spectral hypergraph similarity is developed to minimize the intrinsic topological distribution of neural circuits, which can significantly improve the accuracy and robustness of the proposed model. Experimental results demonstrate that the proposed model can effectively detect the abnormal neural circuits at different stages of AD, which is helpful for pathological study and early treatment.

Motivation & Objective

• To address the challenge of detecting abnormal neural circuits in Alzheimer’s disease from multimodal brain networks (fMRI and DTI), which are often overlooked in traditional methods.
• To develop a deep learning framework that can decouple brain networks into AD-critical circuits and non-influential components, enabling precise pathological analysis.
• To enhance feature representation of AD-related neural circuits through adversarial training and a novel sparse capacity loss that preserves intrinsic topological structure.
• To provide a robust, explainable model for early diagnosis by identifying key neural circuits linked to AD progression across stages.

Proposed method

• A decoupling module is designed to split a multimodal brain network into sparse subgraphs (representing AD-critical neural circuits) and a supplementary graph (negligible influence on AD).
• The generator reconstructs the original brain network from the decoupled components and a latent space, improving feature expression and stability of the decoupling process.
• The discriminator is trained adversarially to distinguish real brain networks from reconstructed ones, enhancing the model’s ability to learn AD-relevant features.
• A hypergraph-based analytic module with hyperedge neurons is used to identify and represent detected neural circuits from the decoupled subgraphs.
• A novel sparse capacity loss is introduced based on spatial-spectral hypergraph similarity to minimize topological differences between neural circuits, improving robustness and accuracy.
• The model leverages graph convolutional networks (GCNs) to extract complementary topological features from functional (rs-fMRI) and structural (DTI) connectivity data.

Experimental results

Research questions

• RQ1 Can a deep generative model effectively decouple multimodal brain networks into AD-critical neural circuits and non-influential components?
• RQ2 How can adversarial training and a novel loss function enhance the detection of abnormal neural circuits in Alzheimer’s disease?
• RQ3 To what extent does the sparse capacity loss improve the topological fidelity and classification performance of detected neural circuits?
• RQ4 Can the reconstructed brain networks derived from the decoupled components outperform priori brain networks in classifying AD stages?

Key findings

• DecGAN achieved 85.18% accuracy, 93.75% sensitivity, and 72.72% F1-score in classifying Alzheimer’s disease (AD) versus normal control (NC), outperforming priori brain networks.
• The reconstruction-based brain networks yielded 86.20% accuracy and 88.88% F1-score in classifying NC vs. AD using GCN, significantly outperforming the priori networks (82.75% accuracy, 83.87% F1-score).
• The sparse capacity loss improved model robustness by minimizing intrinsic topological differences between neural circuits, enhancing generalization across AD stages.
• The detected neural circuits were highly correlated with known AD-affected regions, including the limbic lobe, temporal lobe, and parietal lobe, as validated by clinical literature.
• The model successfully identified abnormal circuits at different AD stages (NC, EMCI, LMCI, AD), supporting its utility in early pathological detection.
• Visualization of reconstructed structural connectivity (weighted adjacency matrices) confirmed that DecGAN preserves meaningful network topology, with clearer AD-related patterns than priori networks.