- P-BASE framework uses two complementary brain atlases to learn dynamic spatiotemporal patterns and long-distance dependencies from rs-fMRI.
- Dual-branch encoder processes sliding-window dynamic FCNs: ST-GCN captures local fluctuations, Transformer branch models long-range relations, Sero readout aggregates node features.
- On the ABIDE dataset P-BASE outperforms several state of the art methods for ASD diagnosis using multi-scale feature fusion for classification.
Neural Netw. 2026 Nov 11;206(Pt C):109898. doi: 10.1016/j.neunet.2026.109898. Online ahead of print.
ABSTRACT
Resting-state functional magnetic resonance imaging (rs-fMRI) is widely used along with machine learning and deep learning approaches to identify neurological conditions such as autism spectrum disorder. These approaches commonly utilize functional connectivity networks (FCNs), which are obtained from blood-oxygen-level-dependent (BOLD) temporal data, to characterize relationships among brain regions of interest (ROIs). However, the brain is a highly dynamic system and capturing its complex spatiotemporal characteristics remains challenging. Graph neural networks such as ST-GCN have shown promise in modeling local spatiotemporal fluctuations in functional connectivity. Moreover, the Transformer has demonstrated significant advantages in modeling long-distance dependencies. Although single-atlas approaches offer a simplified summary of brain organization, they may not fully represent its complex architecture. To address these limitations, we propose a Parallel Bi-Atlas Spatio-temporal Encoder (P-BASE) framework. Using rs-fMRI data, we segmented the brain based on two distinct atlases to learn both dynamic spatiotemporal patterns and long-distance dependencies for the analysis and diagnosis of brain disease. The P-BASE framework comprises four main components: (1) constructing dynamic FCNs from both atlases using sliding windows, (2) processing them through a dual-branch local-global encoder to capture spatiotemporal features, (3) aggregating node features using the Sero readout function, and (4) performing multi-scale feature fusion for classification. Using the ABIDE dataset, our model demonstrates performance superior to several state-of-the-art methods.
PMID:43038664 | DOI:10.1016/j.neunet.2026.109898
Share Evidence Blueprint
Save to Google Notes

Search Google Scholar
⭐ My Revision List

