INTRODUCTION

Movement of the head during scans causes undesirable temporal fluctuations (i.e., noise), which make the identification of the effects that are truly related to the underlying neural activities (i.e., signal) difficult.¹. Recently, ICA-based denoising technique has widely been used as a powerful tool for removing spurious motion-related signal from the data². Furthermore, the framewise displacement (FD) variable, which measures movement of any given frame relative to the previous frame, has been used to regress out the head motion artifacts^3,4 and to standardize a cut-off threshold for excluding high motion-related fMRI datasets^5,6. However, the relationship between FD and ICA analysis has not been much elucidated. By investigating the influence of head motion on the output of ICA-based denoising, the present study aims to clarify their relationships.

METHODS

Twenty-four right-handed young healthy subjects (13 males and 11 females, aged 20 - 25 years), who had no known history of neuropsychiatric disorders or substance abuse, participated in this study. All participants signed written informed consents under an IRB-approved protocol.
In order to increase the possibility for observation of head movements during scans, we used motor tasks. Block-designed fMRI was used. Each fMRI run lasted for 6min30sec, consisting of 7 rest blocks alternatively presented with 6 task blocks. The subjects performed left hand grip motor task at 1Hz frequency during task blocks. Two consecutive fMRI runs of the same task condition were collected from all subjects (test, re-test paradigm). Scans were performed with a 7T whole-body scanner (MAGNETOM 7T, Siemens Healthineers, Erlangen, Germany) using a single-channel transmit and 32-channel receive head coil (Nova Medical, MA, USA). For functional volumes, multiband EPI sequence developed at CMRR⁷ with the following parameters was acquired: TR/TE = 1000/22 ms; FA = 45°; Multiband =5, GRAPPA = 2, spatial resolution = isotropic 1.6 mm. After acquiring the functional images, a whole-brain MP2RAGE anatomical images were acquired (TR/TE/TI1/TI2 = 6000/2.9/800/2700 ms; FA1/FA2 = 4/5°, spatial resolution = isotropic 0.7 mm).
The fMRI dataset was preprocessed by using a default preprocessing pipeline of HCP pipeline protocol (https://github.com/Washington-University/HCPpipelines). Motion correction of EPI images was processed with spatial realignment via MCFLIRT toolbox which is implemented within the default HCP pipeline protocol. The value of absolute root mean squared (RMS) movement of each volume, which describes motion of that volume with respect to the middle volume in time-series, was extracted from MCFLIRT report. The mean motion of an fMRI run is defined as the mean absolute RMS movement of all volumes of that scan. The number of fMRI volumes in which head motion was more than 0.2 mm and 0.5 mm different from adjacent volumes was also quantified⁴. Motion related fluctuation and different sources of artifacts were removed from the images using FMRIB’s ICA-based X-noiseifier (FIX) denoising analysis⁸. The output results of ICA-FIX analysis were re-confirmed by visual inspection of each ICA component⁹ from 2 researchers in the field.

RESULTS

For each fMRI run, the average number of total ICA components was 116.2 (SD = 29) and the average numbers of signal/noise/unknown components were 37.8/62.6/15.8 (32.5/53.9/13.6 %), respectively. These results were re-confirmed by manual labelling, which showed that variance was less than 10% for each component classification (data are not shown).
The total number of ICA components showed a strong correlation with mean motion (Figure 1) and FD (Figure 2). Increase in mean motion resulted in increase and decrease in explained variance of noise and signal components, respectively (Figure 3). The percentage of volumes with FD larger than 0.2mm significantly increased and nearly reached 20% where datasets are occasionally discarded, when the number of ICA components exceeded the value of mean + 2SD (Figure 4).

DISCUSSION

Head movements are well-established as one of the reasons for undesirable fluctuations in fMRI study. The acquisition of BOLD signal depends on precise spatial and temporal placement of magnetic gradients. Movement of the head during scans not only shifts the position of the brain in space but also disrupts the establishment of temporal readout of the BOLD signal^1,10 resulting in spurious motion-related signal in the data. Therefore, evaluating and correcting head motion of an fMRI dataset is crucial. The fact that the number of total ICA components has a strong correlation with both mean motion and FD variables suggested that ICA decomposition can detect the spatial changes of the head throughout the fMRI scan, and the output results (i.e., the number of ICA components) indicated how much the subject moved during the scan.

CONCLUSION

The number of total ICA components from ICA-based denoising analysis can be used as an index for evaluating and detecting high motion-related fMRI datasets.

Acknowledgements

A research grant from Siemens Healthcare K.K., Japan. We also thank Yuta Urushibata from Siemens Healthcare K.K., Japan, for the helpful discussion and data analysis.