Brain State Decoding Based on fMRI Using Semisupervised Sparse Representation Classifications.

Area of Science:

• Neuroimaging
• Machine Learning
• Brain-Computer Interfaces

Background:

• Multivariate classification is crucial for decoding brain states from functional magnetic resonance imaging (fMRI) data.
• Training robust supervised classifiers for fMRI is challenging due to data variability and collection limitations.
• Sparse Representation Classifier (SRC) shows promise but is underutilized in fMRI-based decoding.

Purpose of the Study:

• To enhance the Sparse Representation Classifier (SRC) for effective fMRI-based brain state decoding by incorporating unlabeled testing samples.
• To develop a semisupervised-learning SRC method (semiSRC-AVE) that leverages average coefficients and selective training data updates.

Main Methods:

• Proposed a semisupervised-learning SRC with an average coefficient (semiSRC-AVE) method.
• Utilized the average coefficient of each class for classification, differing from traditional reconstruction error methods.
• Implemented selective updating of the training dataset with high-confidence new testing data.

Main Results:

• Simulated and real fMRI experiments demonstrated the feasibility and robustness of the semiSRC-AVE method.
• SemiSRC-AVE significantly outperformed the supervised learning SRC with an average coefficient (SRC-AVE) method.
• The proposed method showed superior performance compared to three other semisupervised learning techniques.

Conclusions:

• The semiSRC-AVE method offers a significant advancement for fMRI-based brain state decoding.
• This semisupervised approach effectively addresses the challenges of limited labeled fMRI data.
• SemiSRC-AVE presents a robust and high-performing alternative to existing classification techniques in neuroimaging.