MPP-GNN: Subject-Adaptive Community Detection for fMRI-Based Alzheimer's Disease Classification
MPP-GNN introduces a Meta Probabilistic Pooling GNN that performs subject-adaptive community detection for fMRI-based Alzheimer's disease classification, addressing the limitation of preset functional modules across subjects. The model frames its task as a coupled, bilevel optimization: hierarchically discovering subject-specific brain modules and then using those modules as an explicit prior to guide edge refinement and representation learning. MPP-GNN achieves the highest AUC on two public AD
Analysis
TL;DR
- MPP-GNN introduces a Meta Probabilistic Pooling GNN that performs subject-adaptive community detection for fMRI-based Alzheimer's disease classification, addressing the limitation of preset functional modules across subjects.
- The model frames its task as a coupled, bilevel optimization: hierarchically discovering subject-specific brain modules and then using those modules as an explicit prior to guide edge refinement and representation learning.
- MPP-GNN achieves the highest AUC on two public AD classification datasets compared to established baselines.
- Analysis reveals significant alignment with the canonical Yeo brain atlas and uncovers a network-level dedifferentiation pattern associated with Alzheimer's disease.
Why It Matters
This work bridges a critical gap in neuroimaging AI by moving beyond one-size-fits-all graph partitioning to capture inter-subject variability in brain functional organization. For AI practitioners working in medical imaging and graph-based representation learning, it demonstrates how bilevel optimization and probabilistic pooling can produce both high classification performance and neuroscientifically interpretable results. The findings also offer clinicians a more personalized framework for AD detection using fMRI data.
Technical Details
- Architecture: Meta Probabilistic Pooling GNN (MPP-GNN) that combines hierarchical adaptive graph partitioning with probabilistic pooling to discover subject-specific functional modules from fMRI data.
- Bilevel Optimization: The model is framed as a coupled bilevel optimization problem where the upper level discovers subject-specific brain modules via adaptive graph partitioning, and the lower level uses those modules as an explicit prior to guide edge refinement and node representation learning.
- Datasets: Validated on two public fMRI datasets for Alzheimer's disease classification.
- Evaluation: Achieves state-of-the-art AUC performance compared to established baselines on both datasets.
- Neuroscientific Validation: Demonstrates significant alignment with the Yeo brain atlas's canonical functional-network organization and identifies a network-level dedifferentiation pattern characteristic of AD.
Industry Insight
- The subject-adaptive approach challenges the prevailing paradigm of fixed graph partitions in neuroimaging GNNs, suggesting that personalizing brain network structure per subject could be a generalizable strategy for other neurological disorder classifications.
- The bilevel optimization framework offers a reusable template for integrating domain-specific priors (e.g., anatomical atlases) into graph representation learning, which could be extended to other biomedical domains beyond neuroimaging.
- The discovery of network-level dedifferentiation in AD using MPP-GNN provides a quantifiable biomarker that could inform clinical trial design and disease progression monitoring, highlighting the value of interpretable AI in precision medicine.
Disclaimer: The above content is generated by AI and is for reference only.