TL;DR
This paper introduces BrainNetMLP, a novel Multi-Layer Perceptron (MLP) based architecture for functional brain network classification. It achieves state-of-the-art accuracy and significantly improves efficiency by jointly leveraging spatial connectivity and spectral dynamics, offering a compelling alternative to complex Transformer models.
What Problem Does This Paper Address?
Understanding the human brain's functionality and diagnosing neurological disorders often relies on analyzing functional brain networks derived from neuroimaging data like rs-fMRI. These networks are typically represented as graphs where brain regions are nodes and their functional connectivity are edges. While recent deep learning models, particularly Transformers, have shown promise, their increasing complexity often leads to diminishing performance gains and hinders real-world deployment. The authors identify a gap between the escalating complexity of existing models and their practical applicability, seeking a more efficient yet effective approach.
Methods
The core of BrainNetMLP is its dual-branch architecture:
- Spatial Connectivity Mixer (SCMixer): This branch processes the functional connectivity (FC) matrix. Instead of treating it like an image or text, it leverages the symmetry of the FC matrix to extract only the lower triangular part, effectively halving the input size. This reduced feature vector is then fed into a standard MLP to learn spatial connectivity patterns.
- Spectrum ROIs Mixer (SRMixer): This branch focuses on the temporal dynamics of brain activity. It first applies a Fast Fourier Transform (FFT) to the time series data of each Region of Interest (ROI) to obtain its spectral representation. A low-pass filter is then used to remove noise, and an MLP processes the filtered spectrum to capture spectral dynamics.
These two branches are then fused, and a simple non-linear projection layer predicts the classification label. To further enhance efficiency, the authors propose an Edge Degree Guided Pruning (EDGP) technique. This method prunes parameters in the SCMixer based on both their learned weights and the topological importance of the corresponding connections in the brain network, aiming for more biologically plausible and effective pruning.
Datasets
The proposed BrainNetMLP model was evaluated on two publicly available and widely used datasets:
- Human Connectome Project (HCP)
- Autism Brain Imaging Data Exchange (ABIDE)
Key Insights and Ideas
- MLPs are not obsolete: The paper demonstrates that even simple MLPs, when thoughtfully applied and combined with domain-specific insights (like FC matrix symmetry and spectral analysis), can rival or surpass complex, state-of-the-art models like Transformers.
- Efficiency matters for real-world impact: The authors highlight that the pursuit of accuracy shouldn't come at the cost of prohibitive computational complexity, especially for large-scale brain networks. BrainNetMLP's efficiency is a key selling point for practical applications.
- Dual-branch approach is powerful: Combining spatial connectivity with spectral dynamics provides a more comprehensive understanding of brain function, leading to improved classification performance.
- Topology-aware pruning is beneficial: The EDGP method shows that incorporating network structure into pruning can lead to more effective parameter reduction and potentially more interpretable models.
Key Results
- State-of-the-art Accuracy: BrainNetMLP achieved state-of-the-art performance on both the ABIDE and HCP datasets, outperforming existing Transformer, GNN, and CNN models. For instance, on the HCP dataset, it yielded a 2.1% improvement in accuracy and a 1.2% increase in AUCROC compared to the second-best model.
- Significant Efficiency Gains: BrainNetMLP requires substantially fewer FLOPs (computational operations) and parameters compared to Transformer models. The paper claims a reduction of "more than one order of magnitude" in computational complexity.
- Scalability: The model demonstrates better scalability with increasing network size compared to Transformers.
- Faster Inference: BrainNetMLP offers significantly faster inference times, reportedly consuming only one-third to half the time of other methods for classifying a functional brain network.
- Pruning Effectiveness: The EDGP method, particularly at a 50% pruning ratio, showed improved performance on the ABIDE dataset, suggesting the presence of spurious connections that the EDGP can help identify.
Limitations and Future Work
- Pruning on HCP: The performance gain from EDGP was less pronounced on the HCP dataset, possibly due to its smaller number of ROIs (22) compared to ABIDE. This suggests that the benefits of EDGP might be more significant for larger and potentially noisier brain networks.
- Ablation Study Insights: While the full model outperforms its individual components (SCMixer or SRMixer alone), the ablation study indicates that each component contributes uniquely, with SCMixer and SRMixer individually excelling at specificity and sensitivity, respectively. This suggests potential for further optimization by fine-tuning the interaction between these branches.
- Future Work: The authors implicitly suggest that further exploration of MLP-based architectures and advanced pruning techniques could lead to even more efficient and effective brain network analysis tools. The availability of the code also facilitates further research in this direction.
Summary updated