An Adaptively Weighted Averaging Method for Regional Time Series Extraction of fMRI-Based Brain Decoding

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Brain decoding using functional magnetic resonance imaging (fMRI) is crucial for understanding cognitive functions.
• Traditional fMRI analysis averages voxel data within brain regions, losing spatial information.
• Existing methods for cognitive state classification using fMRI face limitations in capturing detailed brain activity patterns.

Purpose of the Study:

• To develop an improved method for brain decoding that preserves spatial information within brain regions.
• To enhance the accuracy and robustness of cognitive state classification from fMRI data.
• To provide a more refined approach for analyzing functional brain activity.

Main Methods:

• A fully connected neural network was developed and jointly trained with a brain decoder.
• The method employs an adaptively weighted average across voxels within each brain region.
• Evaluations were conducted using cognitive state decoding, manifold learning, and interpretability analysis on the Human Connectome Project (HCP) dataset.

Main Results:

• Cognitive state decoding accuracy increased by up to 5% compared to traditional methods.
• Stable accuracy improvements were observed across various time window sizes, resampling sizes, and training data sizes.
• Manifold learning demonstrated enhanced separability of cognitive states and reduced subject-specific information.

Conclusions:

• The proposed method significantly improves fMRI-based cognitive state decoding.
• This approach offers a more nuanced analysis of brain activity by preserving spatial information.
• The findings suggest a potential advancement for the standard fMRI processing pipeline.