Smart Independent Component Analysis (SMART ICA): An Innovative Method for Extracting Accurate Brain Functional Networks

Background Introduction

In brain science research, Functional Networks (FNs) show great potential for understanding human brain function by exploring the integration and interaction relationships between different brain regions. Functional magnetic resonance imaging (fMRI) is an important tool that reveals functional connections between different brain regions by observing changes in blood oxygen level-dependent signals during brain activity. Independent Component Analysis (ICA) is a commonly used data-driven method widely applied to estimate functional networks from fMRI data. However, ICA methods face challenges in determining the optimal model order (i.e., the number of components), which leads to doubts about the reliability of functional network estimation results. Therefore, developing reliable brain functional network analysis methods is particularly important for maximizing the robustness and universality of research results. This paper proposes a method called Smart Independent Component Analysis (SMART ICA), which automatically extracts reliable functional networks from ICA results of multiple model orders through a split-merge assisted method, and verifies its effectiveness in fMRI data analysis of multiple subjects.

Research Source

This paper was written by scholars including Xingyu He, Vince D. Calhoun, and Yuhui Du, from the School of Computer and Information Technology at Shanxi University, the GSU-Georgia Tech-Emory Center for Advanced Brain Imaging and Data Science, and the Center for Excellence in Brain Science and Intelligence Technology of the Chinese Academy of Sciences. The paper was published online on March 15, 2024, and in print in July in “Neuroscience Bulletin”.

Research Methods

Research Process

The overall analysis process of the SMART ICA method can be divided into four main steps:

1. ICA runs with different model orders: First, apply ICA with different model orders to fMRI data of multiple subjects to obtain initial group-level independent components (ICs).

2. Split-merge assisted clustering: Cluster these initial group-level ICs to obtain reliable group-level ICs while establishing association relationships between ICs of different model orders.

3. Remove group-level artifact ICs: Remove artifact ICs and retain reliable group-level functional networks (FNs).

4. Estimate subject-specific functional networks using group information guided ICA (GIG-ICA): Estimate subject-specific FNs and time courses (TCs) based on each subject’s fMRI data using the GIG-ICA method.

Special Techniques and Algorithms

1. ICA run: For p subjects’ data (Xp), first perform principal component analysis (PCA) for dimensionality reduction, then combine all subjects’ data and perform a second PCA to further reduce dimensionality to obtain matrix H. Then apply the improved infomax algorithm for ICA decomposition to obtain group-level ICs of different model orders.

2. Split-merge assisted clustering: Construct a graph structure of all initial group-level ICs and simplify it into a tree structure using graph simplification techniques to guide split and merge operations. Through recursively performing split and merge operations until all tree structures stabilize, achieve automated and robust clustering to integrate functional network information under different model orders.

3. Artifact detection and removal: In simulated data, detect artifacts by measuring the smoothness of ICs and retain ICs with better smoothness. In real data, manually remove artifact ICs, such as those mainly activated in white matter and cerebrospinal fluid.

4. GIG-ICA method: The GIG-ICA method shows better performance by optimizing the independence of subject-specific functional networks and is the preferred method for estimating individual FNs and related TCs.

Experiments and Validation

1. Simulated data validation: Experiments were conducted using two sets of simulated data, each including 100 subjects, to evaluate the effectiveness of the SMART ICA method. Each set of simulated data had predetermined spatial mapping (SM) patterns. Experiments showed that the subject-specific functional networks estimated by the SMART ICA method had high matching with the true SMs.

2. fMRI data validation: Experiments were conducted using data from two age-matched healthy groups from the UK Biobank project. Each group included 975 subjects. The results showed that the SMART ICA method has high consistency in extracting group-level and individual-level functional networks and can identify subtle functional changes with age.

Research Results

1. Simulated data results: Experimental results from two sets of simulated data showed that SMART ICA can effectively extract reliable group-level functional networks from multiple model orders and accurately estimate subject-specific functional networks. The similarity between each subject-specific functional network and the true SM exceeded 0.9, verifying the effectiveness and robustness of the method.

2. fMRI data results: Experiments on real fMRI data showed that SMART ICA can extract highly consistent group-level functional networks in two independent healthy groups and identify functional changes with age. The matching similarity of 24 and 25 reliable group-level functional networks under low model orders exceeded 0.9, and the matching similarity of 74 and 69 functional networks under high model orders was also high, demonstrating the high robustness of the method.

Research Conclusions and Significance

1. Scientific value: The SMART ICA method automatically extracts reliable functional networks from ICA results of multiple model orders without presetting the model order, significantly enhancing the application potential of ICA in brain functional network analysis.

2. Application value: The provided multi-scale functional network templates provide important benchmarks for future research, contributing to the standardization and unification of fMRI data analysis methods.

3. Important findings: The method can identify age-related changes in brain functional networks and demonstrates the trend of weakening brain functional connectivity strength with increasing age.

Research Highlights

1. Automatic split-merge clustering technique: The proposed split-merge assisted clustering method not only efficiently clusters ICs but also captures the association relationships between results of different model orders.

2. High robustness and universality: The method demonstrates high robustness and good universality in both simulated data and real fMRI data, suitable for different brain functional network analysis studies.

3. Multi-scale functional network templates: The provided standardized small-scale and large-scale functional network templates contribute to future large-sample analyses, enhancing the robustness of research results.

The Smart Independent Component Analysis (SMART ICA) method has great application potential in brain functional network research, indicating an important advancement in the field of fMRI data analysis.