Reduction of Motion Artifacts and Noise Using Independent Component Analysis in Task-Based Functional MRI for Preoperative Planning in Patients with Brain Tumor

Discussion

Based on the previously reported success of ICA in improving data quality in both task-based and resting-state fMRI studies, our goal was to investigate its utility in preoperative planning for patients undergoing resection for gliomas. We found improvement in statistical significance and decreased noise, resulting in fewer false-positives and, perhaps more important, an increase in true-positives, similar to findings in studies in patients without tumors.^20,27,28 We also found that a proportion of studies, otherwise nondiagnostic when processed with traditional methods, were diagnostic after ICA denoising.

As previously discussed, underlying motion-related artifacts, physiologic noise, and scanner-related noise, which are ubiquitous in fMRI, have serious detrimental effects on fMRI interpretation. The first of these, subject motion, is becoming an increasingly recognized cause of fMRI misinterpretation. For example, a prominent theory in autism that touted decreased long-range connections was eventually shown to be related to increased head motion in the autistic group relative to controls.^12,29 For any user of fMRI, it is essential to have a thorough understanding of motion and related artifacts. An important tenet of high-quality fMRI is homogeneity of the B0 magnetic field. When an object is introduced into the bore of the scanner, the field becomes distorted and is corrected for with a shimming procedure. However, if the object moves once in the bore, this shim effect is lost and the susceptibility profile of the object changes. This change can result in drastic changes in signal intensity and geometric distortion that cannot be corrected by typical realignment procedures.^11,13 Shifts as small as 2° can potentially create large areas of false-positive activation regardless of intensive thresholding.¹³

A second major consideration regarding patient motion in fMRI is signal change as a result of spin-history artifacts. Given the typical short TRs in fMRI, the brain achieves a steady-state of incomplete longitudinal magnetization recovery after a few TRs. This recovery is slightly out of phase in any individual section. Once the head moves, these voxels will be shifted into a different plane, resulting in either a lower- or higher-than-expected amount of recovery in sections obtained prior or subsequent to the reference section, respectively. This unexpected change in magnetization recovery can have a substantial effect on signal intensity.¹¹ While abrupt motion with a large amount of displacement can have noticeable effects on the signal intensity, minimal amounts of motion (<1-mm displacement) can have a substantial impact on the data if the motion is frequent. Because this movement is low-amplitude, one cannot expect to see significant outliers in mean signal intensity across volumes. Spin-history artifacts are best assessed by DVARS, which assesses the volume-to-volume signal change.²⁵ In our data, we found this frequent, low-amplitude motion (<1 mm) effect to be the most common source of nondiagnostic studies.

Third, when signal fluctuations from artifacts, such as the previously mentioned spin-history artifacts and susceptibility changes, create a time-series similar to the experimental design, commonly referred to as “task-correlated motion,” the artifactual component is generally inseparable from task-related signal change.^8,28,30 Our results show that task-correlated motion has a strong prevalence in nondiagnostic studies, including studies that are not even improved by ICA denoising. Although neurovascular uncoupling could potentially be the source of these nondiagnostic studies, the prevalence of task-correlated motion suggests that this may also be an etiology. Given the impact of TC on study outcomes, the assessment of this parameter should be a routine part of fMRI data quality analysis.

Physiologic noise is also a major source of variance in BOLD fMRI data and is primarily related to fluctuations in basal metabolism, cardiac and respiratory effects, and subtle motion effects from brain pulsation. In fact, it has been shown that physiologic noise at a higher field strength (3T) is the dominant source of noise in fMRI and even counteracts much of the gain in signal-to-noise ratio and contrast-to-noise ratio when moving to higher magnetic field strengths.^9,31 Due to the typically short TRs used in fMRI, the frequency of these artifacts results in their aliasing into the task-related signal.⁷ Because this noise is structured, nonwhite noise, it does not meet the statistical assumption that errors are independent and identically distributed; therefore, it will have a significant impact on most statistical modeling.³² Similarly, scanner-related noise, such as thermal noise, drift, and imperfections in the coil, radiofrequency, gradient, and shim subsystems and various other minor contributors, produces similar effects.⁹ Our study revealed a large decrease in false-positives with use of ICA versus other methods; however, the ICA decrease in false-positives did not correlate with any motion parameters. This finding could suggest that removal of physiologic and scanner noise is also a major contributor to improvement seen with ICA versus other methods.

Many attempts have been made to reduce artifacts in fMRI; however, they remain extremely problematic. There are generally 2 types of artifact-correction systems, hypothesis-driven and data-driven. Hypothesis-driven systems include prospective motion-correction systems, such as the use of navigator sequences,^33,34 optical motion tracking,^16,18,35 or methods of measuring physiologic parameters during the scan (ie, heart rate, respiratory rate, and so forth) that are then used to apply filters to the data or used as nuisance regressors.^7,14 Unfortunately, these methods have been difficult to implement due to inaccurate cardiac or respiratory peak detection, variability in results with differing TRs, lag in the motion data, and increased radiofrequency pulses resulting in decreased temporal resolution. Furthermore, no single method accounts for all sources of noise present in the data.^5,7,32,36

Data-driven approaches, such as removal or regression of volumes with signal changes above a threshold (motion scrubbing) or regression of motion parameters, do not account for all sources of artifacts and have limited effect.^5,25 The limited effect on quality improvement is evident when volumes are discarded on the basis of motion parameters alone, because motion artifacts can have serious detrimental effects, even with minor head motion (<1 mm).⁸ Thus, more thorough parameters are needed when using motion scrubbing, such as DVARS or frame-wise displacement, which can detect very subtle artifactual changes in signal.²⁵ Furthermore, motion scrubbing results in decreased statistical power, particularly in data with numerous motion-affected volumes, while images with detrimental artifacts that do not meet the threshold for rejection still exist. Regression of the motion parameters can also have deleterious effects if the motion is correlated with the task and therefore must be used with extreme caution in task-based experiments.⁶ Even with these considerations, the overall improvement gained by these methods is minimal.^5,12,25 This outcome is also likely due, in part, to the lack of accounting for physiologic and scanner-related noise.

There has been an increasing interest in the use of ICA to denoise fMRI data to improve the sensitivity and specificity of results. The methodology behind ICA is beyond the scope of this article (see Stone³⁷ for an excellent introduction), but in short, ICA deconstructs an fMRI time-series into unique components (each having a spatial map and the corresponding time course), which are maximally spatially independent.^19,22,28,37 These components can then be classified as either artifactual/noise versus task-related activity (see Kelly et al,²⁶ for an excellent review). In addition to manual identification, numerous methods of automated classification have been described, primarily by using machine-learning algorithms, but a discussion of these methods is beyond the scope of this article.^{38⇓⇓⇓⇓–43} The artifactual components consist of a variety of physiologic processes, motion-related artifacts, scanner-related noise, aliasing, and so forth. Once identified, these artifactual components can be removed from the dataset before statistical analysis, such as with the General Linear Model. The improved statistical modeling results in increased sensitivity and specificity, with reduction of false-positives and increased true-positives (Fig 3).^20,27,28

• Download figure
• Open in new tab
• Download powerpoint

Fig 3.

Processed fMRI data from 3 different subjects. Each case shows ICA-denoised data in blue, motion-scrubbed data in red, and overlapping areas of ICA-denoised data and motion-scrubbed data in green. The first case (A) is a motor finger task in a patient with a left parietal glioma. The primary motor cortex for the right finger (arrow) is only seen after ICA denoising. Likewise, the supplementary motor area (arrowhead) shows a slight increase in statistical significance. The second subject (B) is undergoing a motor face task with severe task-correlated motion and severe DVARS showing no major change in the primary motor face cortex (arrow); however, the number of motion-related false-positives (noise) is markedly reduced. The third case (C) is a semantic decision task in which no meaningful activation is present on the motion-scrubbed data. Expected areas of activation in the anterior and posterior language areas are clearly present after ICA denoising.

Previous studies evaluating the use of ICA in task-based presurgical fMRI have yielded mixed results.^21,44,45 While these studies were the first to explore ICA in the presurgical setting, they relied on the extraction of a single component from the ICA analysis results that best modeled the task. This approach can be problematic because task-related activation can often be distributed across multiple components.⁴⁶ An approach rendering the most accurate and reliable results is to remove those components reflecting structured or random noise during preprocessing and then to process the denoised dataset with statistical modeling. Additionally, these studies were performed at 1.5T, in which physiologic noise and susceptibility-related changes are substantially lower than at 3T, so the results may not be directly applicable to higher field strengths.

The results of this study highlight the benefits of ICA denoising in presurgical mapping for tumor resection, even at higher field strengths. The reduction in false-positives related to these nuisance variables also reduces the number of surgical false-positives. Furthermore, this reduction has the effect of providing the surgeon with “cleaner” activation maps. Additionally, we have demonstrated that in some cases, true-positive activation was only identified with the use of ICA denoising, the lack of which could have had detrimental effects on the operation (Fig 4). The appearance of otherwise undetected activation is likely due to improved statistical significance achieved by the removal of overlapping nuisance signal or artifacts in the area of true activation (Fig 5). This effect was most dramatic in patients with a higher incidence of motion-induced artifacts. However, when motion was correlated with the task, the motion-induced signal changes were generally indistinguishable from the task itself and any improvement in data quality with ICA denoising, as well as motion scrubbing, was often negligible. Last, we have shown an overall improvement in statistical significance of activated voxels with the use of ICA denoising.

• Download figure
• Open in new tab
• Download powerpoint

Fig 4.

A and B, Images for a subject undergoing a motor finger task showing areas in which the z score increased (red) after ICA denoising compared with motion scrubbing. The arrow illustrates increased statistical significance in the left finger primary motor cortex adjacent to the tumor. The location of the primary motor cortex for the left finger was confirmed surgically and is shown on the intraoperative map (C), where the crosshairs correspond to primary finger motor cortex. The white “blobs” in C show the thresholded maps for the realignment-only data where there is no activation in the area of the primary motor cortex.

• Download figure
• Open in new tab
• Download powerpoint

Fig 5.

Signal intensity with time from 1 voxel is shown before (A) and after (B) ICA denoising with the expected hemodynamic response in blue. One can appreciate the amount of overlapping noise from a variety of artifacts resulting in a poor fit with the ideal curve. These noise effects are largely removed after ICA denoising, and the subsequent denoised time course shows a substantially improved fit with the task design.

Notable limitations of the current study include its retrospective nature. Because limited evidence exists on the effects of ICA denoising in the surgical setting, we do not believe that the prospective use of this technique is yet warranted. In this retrospective design, a strong inherent bias is present when assessing whether the results would alter patient treatment or decision-making, and as such, this assessment was not included in the analysis. Additionally, precise intraoperative mapping data were not collected in a structured fashion; this omission limits the ability to directly correlate with surgical data in all patients. Likewise, an inherent limitation of many surgical mapping studies is the inability to test all activation sites from limited craniotomies. To address these limitations, we relied on a double-blind expert review of activation maps to assess activation in expected areas on the basis of the task administered. We believe these limitations do not detract from the objective of this study in identifying data-quality improvement with ICA denoising relative to other standard methods of fMRI data denoising. Future prospective studies should be aimed at evaluating the impact on decision-making and patient outcomes.