Independent component analysis: a reliable alternative to general linear model for task-based fMRI

Abstract

Background: Functional magnetic resonance imaging (fMRI) is a valuable tool for the presurgical evaluation of patients undergoing neurosurgeries. Although many pre-processing steps have been modified according to advances in recent years, statistical analysis has remained largely the same since the first days of fMRI. In this study, we examined the ability of Independent Component Analysis (ICA) to separate the activation of a language task in fMRI, and we compared it with the results of the General Lineal Model (GLM).

Methods: Sixty patients undergoing evaluation for brain surgery due to various brain lesions and/or epilepsy and 20 control subjects completed an fMRI language mapping protocol that included three tasks, resulting in 259 fMRI scans. Depending on brain lesion characteristics, patients were allocated to (1) static/chronic not-expanding lesions (Group 1) and (2) progressive/expanding lesions (Group 2). GLM and ICA statistical maps were evaluated by fMRI experts to assess the performance of each technique.

Results: In the control group, ICA and GLM maps were similar without any superiority of either technique. In Group 1 and Group 2, ICA performed statistically better than GLM, with a p-value of < 0.01801 and < 0.0237, respectively. This indicated that ICA performs as well as GLM when the subjects are able to cooperate well (less movement, good task performance), but ICA could outperform GLM in the patient groups. When both techniques were combined, 240 out of 259 scans produced reliable results, showing that the sensitivity of task-based fMRI can be increased when both techniques are integrated with the clinical setup.

Conclusion: ICA may be slightly more advantageous, compared to GLM, in patients with brain lesions, across the range of pathologies included in our population and independent of symptoms chronicity. Our findings suggest that GLM analysis may be more susceptible to brain activity perturbations induced by a variety of lesions or scanner-induced artifacts due to motion or other factors. In our research, we demonstrated that ICA is able to provide fMRI results that can be used in surgery, taking into account patient and task-wise aspects that differ from those when fMRI is used in research.

Keywords: GLM; ICA; brain mapping; epilepsy; fMRI; language mapping; neuroimaging; presurgical evaluation.

Figure 1

Flowchart of language map scoring for both techniques.

Figure 2

Examples of language maps for each score. Score 0 is not shown as it is a blank/noisy map. Images are in radiological orientation. GLM scoring when the score of ICA analysis was 0, 1, or 2. ICA scoring when the score of GLM analysis was 0, 1, or 2. GLM, General Linear Model. ICA, Independent Component Analysis.

Figure 3

The sum of the scorings for each analysis technique. GLM, General Linear Model; ICA, Independent Component Analysis. As can be observed, GLM presented a motion-related artifact that interfered with the results, making them unreliable. ICA was able to differentiate and split these two signal sources into different components producing a reliable map. Images are in radiological orientation.

Figure 4

Scoring of each analysis technique when the other technique produced unreliable results. (A) GLM scoring when the score of ICA analysis was 0, 1, or 2. (B) ICA scoring when the score of GLM analysis was 0, 1, or 2. GLM, General Linear Model; ICA, Independent Component Analysis. As can be observed, ICA merged in a single component the activation with an artifact source resulting in an unreliable map. The ICA time series is far from the task time series. GLM produced a reliable map showing that the language areas were activated according to the task. Images are in radiological orientation.

Figure 5

Example of a task that scored “1” in GLM analysis while scoring a “5” in ICA analysis. As can be observed, GLM presented a motion-related artifact that interfered with the results, making them unreliable. ICA was able to differentiate and split these two signal sources into different components, producing a reliable map. Images are in radiological orientation.

Figure 6

Example of a task that scored “1” in ICA analysis while scoring a “4” in GLM analysis. As can be observed, ICA merged in a single component the activation with an artifact source resulting in an unreliable map. The ICA time series is far from the task time series. GLM produced a reliable map showing that the language areas were activated according to the task. Images are in radiological orientation.