TLDR: NeuroPathNet is a novel framework that models the dynamic evolution of connections between brain functional regions as “path trajectories” using temporal neural networks. It outperforms existing methods in classifying neurological conditions like Alzheimer’s, autism, and cognitive states on fMRI datasets, offering a more detailed and interpretable analysis of brain functional connectivity.
The human brain is an incredibly complex and dynamic system, and understanding how its functional networks evolve over time is crucial for both comprehending cognitive processes and diagnosing neurological conditions. Traditional methods for analyzing brain networks often struggle to capture these temporal changes, especially in the connections between specific functional communities.
To address this challenge, a new framework called NeuroPathNet has been proposed. This innovative model focuses on characterizing the dynamic behavior of connection pathways between different functional regions of the brain. Instead of viewing the brain as a static system, NeuroPathNet treats the functional connections between each pair of brain communities as independent “path trajectories.” Each of these trajectories is then modeled using a temporal neural network to learn its unique dynamic evolution.
The NeuroPathNet framework begins by taking fMRI (functional Magnetic Resonance Imaging) data as input. fMRI is a powerful tool that helps researchers observe brain activity patterns by detecting blood oxygen level dependent (BOLD) signals. Based on established static partitioning schemes, such as Yeo and Smith ICA, the brain is divided into multiple functional modules. The strength of connections between these modules is then extracted over time, forming dynamic path sequences.
Each path sequence is encoded as a time-series vector, incorporating positional information to maintain temporal order. These vectors are then fed into a specialized Path Transformer, which is designed to model the temporal interactions within each path. Following this, a cross-path attention mechanism integrates all these individual path representations. This allows for a joint analysis of multiple functional pathways and helps generate a comprehensive representation of the whole brain’s activity. Finally, an attention pooling module and a classifier are used to produce the final prediction, which can be used for diagnostic tasks.
The researchers validated NeuroPathNet’s performance on three publicly available fMRI datasets: ADNI (Alzheimer’s Disease Neuroimaging Initiative), ABIDE (Autism Brain Imaging Data Exchange), and HCP (Human Connectome Project). The results consistently showed that NeuroPathNet outperformed existing mainstream methods across various evaluation indicators. For instance, in the ADNI dataset, NeuroPathNet achieved an accuracy of 66.67% in distinguishing normal, mild cognitive impairment, and Alzheimer’s disease, surpassing the next best model. On the ABIDE dataset, it achieved an accuracy of 79.21% in classifying autism spectrum disorder, demonstrating its effectiveness in capturing ASD-related brain network patterns. Similarly, for the HCP dataset, the model showed state-of-the-art performance in cognition classification, with an AUC of 86.26%.
An ablation study further confirmed the importance of each core module within NeuroPathNet, showing that the model’s performance improved with the gradual addition of path modeling, global integration, multi-head attention, and time pooling. The study also explored different brain region partitioning strategies, finding that a finer-grained partitioning like Schaefer-100 (aggregated into 10 functional networks) yielded the best results, suggesting that better spatial granularity helps the model learn interactive features more effectively.
Also Read:
- Decoding Brain Signals into Images with MindHier
- A Geometric Map for Comparing AI and Human Brain Intelligence
This study marks a significant step forward in dynamic graph learning methods for brain network analysis. By providing a structured and interpretable way to model the dynamic interactions between brain regions at a path level, NeuroPathNet offers promising avenues for clinical applications in the diagnosis of neurological diseases. The source code for NeuroPathNet is openly available for further research and development. You can find more details about this research in the full paper available here.
