Hierarchical multi-resolution mesh networks for brain decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• The human brain processes information across multiple frequency bands.
• Functional Magnetic Resonance Imaging (fMRI) data contains rich information when analyzed at various resolutions.

Purpose of the Study:

• To introduce Hierarchical Multi-resolution Mesh Networks (HMMNs) for enhanced brain decoding from fMRI data.
• To represent cognitive processes using multi-resolution brain networks derived from fMRI signals.

Main Methods:

• Decomposition of fMRI signals into frequency subbands using wavelet transform.
• Formation of brain networks at each subband by ensembling local meshes with arc weights estimated via ridge regression.
• Training classifiers using an ensemble learning architecture (fuzzy stacked generalization - FSG) on adjacency matrices from different subbands.

Main Results:

• HMMNs achieved 99% accuracy in discriminating cognitive tasks on the Human Connectome Project task-fMRI dataset across 808 subjects.
• The framework demonstrated superior decoding performance compared to single-classifier approaches or methods using pairwise correlations/average voxel time series.
• Fusion of diverse information using FSG outperformed majority voting for combining subband data.

Conclusions:

• fMRI data during cognitive tasks contain diverse, complementary information within multi-resolution mesh networks.
• HMMNs effectively fuse this multi-resolution information to significantly boost brain decoding performance.
• Mesh networks offer superior representation power for fMRI data compared to traditional methods.