Introduction
Mapping the functional connectome—the complex network of synchronized activity across brain regions—is essential for diagnosing and understanding neurological disorders. However, researchers face a significant challenge in balancing model complexity with biological accuracy. Traditional machine learning methods often fail to capture the non-linear relationships within the brain, while modern Transformer-based models suffer from quadratic computational complexity ($O(N^2)$), making them inefficient for high-resolution data. Furthermore, most deep learning models treat the brain as a generic graph, ignoring fundamental organizational principles.
To address these issues, we introduce the Core-Periphery State-Space Model (CP-SSM). This framework combines the computational efficiency of State Space Models (SSMs) with the biological "Core-Periphery" principle to improve both the accuracy and scalability of functional connectome classification.
Methodology
The CP-SSM framework is built on two innovative pillars designed to optimize how brain connectivity is processed:
1. Selective State Space Models (Mamba)
Rather than utilizing the resource-heavy attention mechanisms found in Transformers, CP-SSM leverages the Mamba architecture. Mamba provides linear computational complexity ($O(N)$), allowing the model to capture long-range dependencies across the brain more efficiently. This reduction in overhead makes the model better suited for the long sequences and large region sets found in modern fMRI scans.
2. Core-Periphery Mixture-of-Experts (CP-MoE)
In standard deep learning, Mixture-of-Experts (MoE) layers use a router to assign data to specific sub-networks. CP-SSM redesigns this router using the Core-Periphery (CP) principle. By partitioning brain nodes into "core" regions (dense, high-traffic hubs) and "periphery" regions (specialized, sparse connectors), the model uses a structural prior to guide its learning process. This bio-inspired inductive bias helps the model focus on the most discriminative connectivity features.
The final pipeline takes a Functional Connectivity (FC) matrix derived from BOLD signals, processes it through cascaded CP-SSM blocks, and utilizes an orthonormal clustering readout for classification.
Experimental Setup and Results
The model was evaluated using two major benchmark fMRI datasets: the ABIDE (Autism Brain Imaging Data Exchange) dataset and the ADNI (Alzheimer’s Disease Neuroimaging Initiative) dataset.
Classification Performance
CP-SSM consistently outperformed traditional machine learning, CNNs, and state-of-the-art Transformer models:
- ABIDE (ASD vs. NC): Achieved an AUROC of 82.4%, surpassing the best Transformer baseline (79.6%).
- ADNI (MCI vs. NC): Achieved an AUROC of 80.7%, a significant improvement over Transformer-based models, which typically performed between 73% and 74%.
Efficiency and Ablation
The transition to an SSM-based architecture allowed for a reduction in parameter count (from 12.2M to 11.6M) while simultaneously increasing accuracy. Ablation studies further demonstrated that the Core-Periphery guidance is vital; removing the CP-MoE component led to a notable drop in performance, proving that biological priors are essential for modeling brain data.
Biological Interpretability
One of the primary advantages of CP-SSM is its ability to identify key discriminative brain regions that align with known neuroscience.
- For Autism Spectrum Disorder (ASD), the model highlighted the Precuneus and Middle Frontal Gyrus, regions associated with social cognition.
- For Mild Cognitive Impairment (MCI), the model flagged the Anterior Cingulate Gyrus and Central Sulcus, which are linked to neurodegeneration.
Limitations and Future Work
While CP-SSM shows great promise, it is sensitive to the "core node rate" hyperparameter, which currently requires manual tuning based on the dataset's sparsity (e.g., 0.2 for ABIDE vs. 0.8 for ADNI). Future work will focus on automating this parameter selection and extending the model to regression tasks, such as predicting disease progression or cognitive scores over time.
Conclusion
By integrating the efficiency of Mamba with the biological reality of core-periphery brain organization, CP-SSM provides a powerful and scalable tool for neurological disorder classification. It demonstrates that bio-inspired constraints can lead to models that are not only more efficient but also more accurate and interpretable for clinical applications.
Summary updated