Neural network-based analysis of MR time series.

Area of Science:

• Neuroimaging
• Data Science
• Medical Imaging

Background:

• Functional magnetic resonance imaging (fMRI) generates complex time-series data.
• Clustering methods are used to analyze fMRI data by grouping similar temporal behaviors.
• Identifying dynamic effects, such as functional activation, within these clusters is a key goal.

Purpose of the Study:

• To introduce a conceptual extension to clustering using self-organizing maps (SOMs) for fMRI data analysis.
• To leverage SOMs for enhanced data visualization and summarization of dynamic effects.
• To explore the application of this method to fMRI and MR mammography.

Main Methods:

• Utilizing self-organizing maps (SOMs) for partitioning fMRI time-series data.
• Employing the two-dimensional projection plane of SOMs to order cluster centers.
• Analyzing the capability of SOMs to visualize and summarize dynamic effects through data partitioning.
• Investigating the formation of 'superclusters' to accommodate varying cluster sizes and populations.

Main Results:

• SOMs provide additional information through the ordered arrangement of cluster centers on a projection plane.
• The method allows for effective data visualization, summarizing dynamic effects via partitioning.
• SOMs can form 'superclusters,' accommodating clusters of different sizes and population densities.
• The approach is demonstrated with applications in fMRI and MR mammography.

Conclusions:

• Self-organizing maps offer a powerful extension to traditional clustering for analyzing complex neuroimaging data.
• This visualization-centric approach aids in identifying and summarizing dynamic patterns, including functional activation.
• The method shows promise for applications beyond fMRI, such as MR mammography.