Ensembling brain regions for brain decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Brain decoding aims to infer cognitive states from brain activity.
• Existing methods often rely on voxel selection or random subspace ensembles.
• Understanding semantic representations in the brain is crucial for decoding.

Purpose of the Study:

• To propose a novel ensemble method for brain decoding using naturally clustered brain regions.
• To leverage the distributed nature of semantic representations for improved classification accuracy.
• To demonstrate the superiority of brain region-based ensembles over random ensembles.

Main Methods:

• Functional MRI (fMRI) data were collected during a semantic category task.
• Hierarchical agglomerative clustering with correlation similarity was used to form voxel clusters.
• Support vector machine (SVM) classifiers were trained on each voxel cluster to predict class-posteriori probabilities.
• A meta-layer SVM was trained on the ensembled probabilities for final cognitive state classification.

Main Results:

• The proposed method achieved higher classification accuracies on the Mitchell dataset for most subjects.
• The brain region-based ensemble approach outperformed state-of-the-art methods using voxel selection.
• Ensembling class-posteriori probabilities from voxel clusters enhanced classification performance.

Conclusions:

• Brain regions offer a natural and effective basis for ensemble learning in brain decoding.
• The proposed method effectively utilizes complementary semantic representations within the brain.
• This approach represents a significant advancement in decoding cognitive states from fMRI data.