Introduction

Several tools can perform ICA on fMRI data (CONN, Melodic in FSL, GIFT, others?). So, why choose GIFT?

Applications Compared

• Not all ICA tools are the same. Here I compare three commonly used ICA tools for fMRI analysis: Melodic, GIFT, and CONN.
• MELODIC is an ICA toolbox in FSL. It was first released in 2001 by Beckman and colleagues, and uses tICA (temporal ICA). It is routinely used to differentiate signal from noise and remove noise components. For example, fMRIprep implements the ICA-AROMA denoising procedure using Melodic ICA.
  • PROS: Implementation is independent of Matlab, and it is good for cleaning up noise in data.
  • CONS: Not as oriented to visualizing and evaluating group-level ICs or running network connectivity analyses.
    
• GIFT is a Matlab toolbox first released in 2004 by Vince Calhoun and colleagues. GIFT focuses squarely on Group ICA analysis of fMRI data: instead of just applying ICA to individual subjects to clean the data, you can apply ICA to a large group of subjects to identify shared patterns. In this way, it is similar to GLM, but really much more informative. GIFT continues to develop increasingly robust and sophisticated approaches to blind source separation (ICA, IVA, and others), including approaches that better handle noise, component overlap, and subgroup and individual differences.
  • PROS: Cutting-edge ICA algorithms, extremely flexible, built to perform analysis and look for group patterns.
  • CONS: Complex choices, analysis-only (no preprocessing pipeline)
    
• CONN, first released in 2011 by Susan Whitfield-Gabrieli, Nieto-Castonon, and colleagues, is a user-friendly toolbox implemented in Matlab. It provides a whole pipeline for processing fMRI data, starting with DICOM conversion and including preprocessing, analysis, and visualization. It implements several approaches to exploring functional connectivity in fMRI. One of those connectivity tools is an ICA implementation that follows GIFT's general methodology but implements only a single ICA algorithm, FastICA, described in Calhoun's original 2001 paper. FastICA is a commonly used ICA implementation, but its performance suffers in the presence of noise and variability between individuals.
  • PROS: Complete processing and analysis pipeline for functional connectivity analyses, with many options. Very user-friendly.
  • CONS: The ICA option only implements FastICA, which may not be the optimal choice.

Watch this ~10-minute video clip

Categories of Blind Source Separation Algorithm

As you could see from the preceding section, ICA is part of a family of related blind source separation techniques. The algorithms fall into several separate categories with different pros and cons, so it is helpful to be aware of these categories and their implications.

Spatial vs Temporal

A major distinction is between spatial and temporal ICA.

GIFT and CONN use spatial ICA (sICA) which optimizes independence over voxels. Spatial ICA is good for fMRI because spatial patterns tend to be more stable than temporal patterns (Calhoun 2012) and spatial ICA can identify more noise-related components (Golestani 2022).

In comparison, temporal ICA optimizes independence over time. Temporal ICA was better than one sICA algorithm (FastICA) at identifying physiological noise like heartbeat (Golestani 2022).

Golestani ends up advocating for a hybrid approach that uses temporal and spatial ICA to clean resting-state data. Interestingly, fMRIprep implements data cleaning with temporal ICA, which might be a good choice before running a spatial ICA algorithm in GIFT. That is, running fmriprep prior to GIFT may provide the sort of hybrid approach advocated by Golestani 2022. Again, the details could make a lot of difference.

In sum, there are different variants of ICA, each with PROS and CONS. One of the most basic distinctions is between temporal ICA and spatial ICA. Both have a role to play. This lesson will explore different spatial ICA algorithms as implemented in GIFT.

Deterministic vs Iterative

Another important distinction amongst spatial ICA algorithms is whether they are deterministic or iterative.
ICA algorithms that produce identical results for each run are deterministic. This includes AMUSE, JADE, ERICA, RADICAL, and SIMBEC. So, you only need to run the algorithm once.

However, many popular ICA algorithms, such as Infomax and FastICA, generate different results for each run, which is generally acknowledged as a drawback for ICA approaches. Such algorithms are said to be stochastic or non-deterministic. And, in order to address the issue of variability, such algorithms should be run iteratively to identify statistically consistent results.

Watch this ~9-minute video clip

Model Order Selection (Stability Analysis)

Model Order Selection algorithms like ICASSO address two related issues:
1) Selecting the correct number of components, and
2) Identifying good components

• All three tools (Melodic, GIFT, and CONN) offer some way to estimate the optimal number of ICs and give you a choice to use that estimate or to specify the number of components you want. However, GIFT provides some excellent tools for identifying good components. This is especially relevant for iterative algorithms.
• ICA separates signals into more and more categories as you increase the number of components. You can represent this as a tree: when you ask for more components, you are asking for the tree to fork its branches.
• If you underestimate the number of ICs, you lose fine-grained distinctions (Abou-Elseoud 2010).
• If you overestimate the number of ICs, the components will not generalize well to other data because the components become unstable, and estimation of task-related activity degrades (Li 2007).

Order selection tries to identify the Goldilocks zone with the appropriate number of stable components.
However, Abou-Elseoud 2010 reported that it is often better to choose a higher number of components because more components provide finer distinctions (up to the point where they become unstable). In addition, you can always follow the branches in the tree backward to determine which ICs would group together if there were fewer components.

ICASSO in GIFT is useful for both selecting the right number of components and identifying good components.

Watch this ~5-minute video clip

ICASSO

• Icasso is a pun on ICA and Picasso because the primary author thought the results looked cubistic.
• Use ICASSO to run non-deterministic ICA iteratively (10 times or so) and reveal stable statistical patterns, that is, compact and isolated components.

Icasso Similarity graphs

• The similarity graph is an excellent way to represent components.
• It gives each IC an identification number.
• The proximity of the components reflects their similarity.
• The IC is characterized by a hull which wraps around the individual estimates.
• The most stable components are isolated from other clusters. They are also compact (the dots are close together and the hull that bundles the dots like a rubberband is therefore small).
• In the video, you'll see 10 black dots for each component because I ran ICASSO 10 times.
• The centrotype is the dot in the blue circle. It has the maximum correlation to other points in the cluster and is the best representative of that component.
• The red color means the within-cluster correlation is greater than 0.9. A fully red component is probably a keeper.
• Components in  pink   have a within-cluster correlation between 0.8 and 0.9. They are suspect, and other factors (like whether it is signal or noise, whether it is task-related, its spatial distribution, etc.) should be considered when deciding whether or not to keep them in the analysis.
• The least stable components are mostly white, indicating a within-cluster correlation of less than 0.8. They are likely to be less compact than the stable components. These are suspect, and should be evaluated by other methods.
• Finally, overlapping components are not isolated, and probably not acceptable.

Watch this ~16-minute video clip

FNC and dFNC

Before we talk about selecting the best algorithms for your analysis goals, consider two common post-ICA analyses you might want to perform. Your choice of ICA algorithms should be influenced by the sorts of subsequent analyses you want to perform, in addition to other factors like the time and computational power available, and how interested you are in diversity (individual and subgroup variation) in your data.

I'll focus on functional connectivity (a.k.a group networks), and dynamic functional connectivity. Both are often used after running ICA analysis in an effort to identify relationships among the components.

What do (FNC) functional network connectivity and dynamic functional network connectivity (dFNC) mean?
Independent components can be correlated with each other! This might seem contradictory, but when we say components are independent, that just means they have different sources, it does not prevent those sources from interacting!

FNC (Functional Network connectivity) provides a single average snapshot of the correlations for the entire time course.
FNC is a good fit for identifying resting state networks.

However, functional relationships between components may be complex. Tracking these sorts of complex and varying interactions between components is the job of dFNC (dynamic functional network connectivity) analysis, which uses a sliding window to identify changes over time (that is, it does functional network connectivity over and over again every few seconds along the timeline).

Watch this ~6-minute video clip

Blind source separation alternatives

• Whereas CONN offers only fastICA, GIFT offers 17 ICA algorithms, and many of these offer variants! How can you choose among them?
• First, to get the most out of GIFT, combine all subgroups, participants and sessions into one big group and run them together so the component numbers will match across participants and sessions.

Choosing an Algorithm

• The algorithms are roughly arranged in order from the earliest algorithms, like Infomax and FastICA, to the newest IVA algorithms.
• Newer algorithms, like IVA, increasingly handle diversity better, that is, they better account for individual and subgroup variation while still identifying patterns in the whole group. In general, however, the newer algorithms are more computationally demanding.

GIFT ICA Algorithms

Details

• Algorithms 13, 14, and 16 are recommended for resting state networks, in part because they handle diversity better.
• GIG-ICA (#16) may be especially good for functional connectivity.
• IVA algorithms (15 and 17) are superior to ICA for capturing subject variability (Bhinge 2019), and the performance of IVA-GL actually improves as group variability increases (Ma 2013)!
• A variant of IVA-L (Adaptively constrained IVA-L-SOS) is especially good for dFNC (Iraji 2021).
• Finally, comparisons of the different algorithms on a number of features have resulted in bad reviews for RADICAL (very slow), EVD, SIMBEC, and ERICA (Sariya 2017).

Watch this ~9.5-minute video clip

Articles

• Abou-Elseoud A, Starck T, Remes J, Nikkinen J, Tervonen O, Kiviniemi V (2010) The effect of model order selection in group PICA. Human Brain Mapping NA.
• Bhinge S, Mowakeaa R, Calhoun VD, Adali T (2019) Extraction of Time-Varying Spatiotemporal Networks Using Parameter-Tuned Constrained IVA. IEEE Trans Med Imaging 38: 1715-1725.
• Calhoun VD, Eichele T, Adali T, Allen EA (2012) Decomposing the brain: components and modes, networks and nodes. Trends in Cognitive Sciences 16: 255-256.
• Golestani AM, Chen JJ (2022) Performance of Temporal and Spatial Independent Component Analysis in Identifying and Removing Low-Frequency Physiological and Motion Effects in Resting-State fMRI. Front Neurosci 16: 867243.
• Iraji A, Faghiri A, Lewis N, Fu Z, Rachakonda S, Calhoun VD (2021) Tools of the trade: estimating time-varying connectivity patterns from fMRI data. Soc Cogn Affect Neurosci 16: 849-874.
• Li Y-O, Adali T, Calhoun VD (2007) Estimating the number of independent components for functional magnetic resonance imaging data. Human Brain Mapping 28: 1251-1266.
• Ma S, Phlypo R, Calhoun VD, Adali T (2013) Capturing group variability using IVA: A simulation study and graph-theoretical analysis. IEEE International Conference on Acoustics, Speech and Signal Processing Proceedings
• Sariya YK, Anand RS (2017) Comparison of separation performance of independent component analysis algorithms for fMRI data. J Integr Neurosci 16: 157-175.
• Soldati, N., Calhoun, V. D., Bruzzone, L., & Jovicich, J. (2013). ICA analysis of fMRI with real-time constraints: an evaluation of fast detection performance as function of algorithms, parameters and a priori conditions. Front Hum Neurosci, 7, 19.
• Wei P, Bao R, Fan Y (2022) Comparing the reliability of different ICA algorithms for fMRI analysis. PLoS One 17: e0270556.

Other

• Dianne Patterson's GIFT Google Doc Tutorial (This includes a glossary in the back.)
• Group ICA/IVA of fMRI Toolbox (GIFT) Manual The GIFT Documentation Team Jan 15, 2020
• ICA_algorithm_table: A nice cheat sheet
• Why tICA? Human Connectome Project Slides

Web pages

• CONN
• fmriprep ICA AROMA
• fMRIprep and GIFT: Examples of fMRIprep outputs usage for group ICA/resting state?
• FSL Melodic
• GIFT Website

Summary

This lesson focused on the differences among ICA algorithms. I introduced Model Order Selection and Icasso which are relevant for some algorithms. I also described functional and dynamic functional connectivity and talked about how the choice of algorithm would affect the success of such post-ICA analyses.

TO DO Checklist

✅ Participate in the Discussion
● fMRI Resources
● fMRI Tools
● Student Supplied Threads: fMRI
● fMRI analysis