Optimal Model-free Approach Based on MDL and CHL for Active Brain Identification in fMRI Data Analysis

Area of Science:

• Neuroimaging
• Machine Learning
• Data Analysis

Background:

• Cluster analysis is vital for uncovering data structures and grouping similar objects.
• In Functional Magnetic Resonance Imaging (fMRI), clustering classifies voxels based on time-course signals to identify hemodynamic responses.

Purpose of the Study:

• To introduce and evaluate a novel unsupervised learning approach using the Enhanced Neural Gas (ENG) algorithm for fMRI data analysis.
• To compare the efficacy of ENG with the traditional Neural Gas (NG) method and Statistical Parametric Mapping (SPM) in fMRI research.

Main Methods:

• The study proposes an unsupervised learning approach utilizing the Enhanced Neural Gas (ENG) algorithm, building upon the Neural Gas (NG) network structure.
• ENG employs a prototype-based clustering strategy for analyzing fMRI data.
• Performance was validated using four indices: Jaccard Coefficient (JC), Receiver Operating Characteristic (ROC), Minimum Description Length (MDL), and Minimum Square Error (MSE).

Main Results:

• ENG demonstrated superior performance compared to NG and SPM on real auditory fMRI data.
• ENG showed robustness against data ordering, initialization variations, and outliers.
• The method effectively determined the optimal cluster number and identified active brain areas.

Conclusions:

• The Enhanced Neural Gas (ENG) technique effectively addresses the limitations of NG in fMRI analysis.
• ENG accurately identifies active brain regions and determines cluster centers using MDL values during network learning.
• This unsupervised learning approach offers a powerful tool for fMRI data exploration and analysis.