Multi-Scale Spatio-Temporal Fusion with Adaptive Brain Topology Learning for fMRI Based Neural Decoding

Area of Science:

• Neuroscience
• Computational Neuroscience
• Brain-Computer Interfaces

Background:

• Neural decoding analyzes brain activity for insights into brain function.
• Spatio-temporal computing is crucial for neural decoding due to brain signal characteristics.
• Existing methods often use static brain topology, missing dynamic interactions and hierarchical organization.

Purpose of the Study:

• To propose a novel framework, MSST-ABTL, for enhanced neural decoding.
• To address limitations of static brain topology in current spatio-temporal decoding methods.
• To improve the understanding of dynamic and hierarchical brain interactions.

Main Methods:

• Developed the Multi-Scale Spatio-Temporal framework with Adaptive Brain Topology Learning (MSST-ABTL).
• Incorporated an Adaptive Brain Topology Learning (ABTL) module to learn dynamic brain topology.
• Integrated an MSST module to capture spatio-temporal associations and multi-scale interpretability.

Main Results:

• MSST-ABTL demonstrated superior performance compared to state-of-the-art methods on the HCP dataset.
• The framework achieved better results across four evaluation metrics.
• Experiments utilized both resting-state and task-related fMRI data.

Conclusions:

• MSST-ABTL significantly advances neural decoding capabilities.
• The framework offers deeper insights into dynamic and hierarchical brain organization.
• This approach holds potential for renewing neuroscientific discoveries.