Adaptive Critical Subgraph Mining for Cognitive Impairment Conversion Prediction with T1-MRI-based Brain Network

TL;DR

This work introduces Brain-SubGNN, a dynamic, graph-based framework that constructs subject-specific brain networks from T1-MRI and adaptively mines critical loop and neighbor subgraphs to predict conversion to cognitive impairment. By combining adaptive node/edge construction, RL-driven subgraph mining, and mutual-information augmentation, the model captures both local and long-range connectivity with explicit subgraph-level interpretability. The approach shows superior performance on MCI conversion prediction and demonstrates feasibility for early diagnosis from NC to MCI, with visualization highlighting known disease-relevant regions. The framework offers a data- and task-driven pathway for robust, interpretable brain-network analysis with potential for clinical deployment.

Abstract

Prediction the conversion to early-stage dementia is critical for mitigating its progression but remains challenging due to subtle cognitive impairments and structural brain changes. Traditional T1-weighted magnetic resonance imaging (T1-MRI) research focus on identifying brain atrophy regions but often fails to address the intricate connectivity between them. This limitation underscores the necessity of focuing on inter-regional connectivity for a comprehensive understand of the brain's complex network. Moreover, there is a pressing demand for methods that adaptively preserve and extract critical information, particularly specialized subgraph mining techniques for brain networks. These are essential for developing high-quality feature representations that reveal critical spatial impacts of structural brain changes and its topology. In this paper, we propose Brain-SubGNN, a novel graph representation network to mine and enhance critical subgraphs based on T1-MRI. This network provides a subgraph-level interpretation, enhancing interpretability and insights for graph analysis. The process begins by extracting node features and a correlation matrix between nodes to construct a task-oriented brain network. Brain-SubGNN then adaptively identifies and enhances critical subgraphs, capturing both loop and neighbor subgraphs. This method reflects the loop topology and local changes, indicative of long-range connections, and maintains local and global brain attributes. Extensive experiments validate the effectiveness and advantages of Brain-SubGNN, demonstrating its potential as a powerful tool for understanding and diagnosing early-stage dementia. Source code is available at https://github.com/Leng-10/Brain-SubGNN.

Figures (8)

• Figure 1: There are some changes in brain through cognitive impairment. (a) critical regions related to atrophy or some other lesion, (b) closed-loop mechanisms in brain self-regulation, (c) local neighbor connectivity.
• Figure 2: The workflow of our proposed method. It assesses brain disorders in three steps: 1) constructing adaptively data-specific and task-specific structural brain networks with both dynamic nodes and dynamic connections; 2) mining critical subgraphs through two Modules, which adaptively mine local loop and neighbor subgraphs of task-oriented, individually heterogeneous and arbitrarily size and shape; 3) encoding loop and neighbor subgraphs, and fuses subgraphs information with global graphs using shallow graph convolution for downstream tasks.
• Figure 3: The workflow of subgraph mining module. For a given center node, 1) the LSRM generates an action $a_t^l$ by policy $\pi^l$ to detect the critical loop subgraph; 2) the NSRM generates an action $a_t^d$ by policy $\pi^d$ to choose the depth of subgraph, then the policy $\pi^w$ generates an action $a_t^w$ to choose the width of subgraph.
• Figure 4: The agents of subgraph mining module.
• Figure 5: Effects of the coefficient of MI module in controlling the contribution of the Global-Aware MI loss.