Purpose

High-field, high-resolution fMRI is opening the door to non-invasive cortical depth-dependent connectivity analyses. However, the depth-dependent variations in baseline physiology [1] (e.g. blood volume, oxygenation) and their corresponding susceptibility to physiological noise [2] can hamper the interpretation of depth-dependent results, especially in resting-state data. The purpose of this study is to examine, quantify, and compare several common methods for physiological noise cleaning, including regression of cardiac and respiratory traces, blood vessel ROI time series, and specific ICA components. We discuss results of these noise cleaning methods with respect to the specific signature of physiological noise at high resolution. Doing so, we are focusing on the case of cortical depth-dependent blood volume signal in human motor cortex.

Methods

Two healthy volunteers were scanned on a 7T Siemens scanner with a 32-channel NOVA Medical head coil at a resolution of 0.75x0.75 mm² with 1.8 mm slice thickness, aligned perpendicular to the cortical surface [3]. Simultaneous physiological monitoring via pulse oximeter and pneumatic respiratory belt was employed. Physiological signals were sampled with a Biopac® system, synchronized with fMRI data acquisition. Respiratory and cardiac signals were sampled at each TR (3 seconds) and later used as regressors. Volunteers performed a 30 second on-off finger tapping task, for a 720 seconds. Motion corrected image time series were corrected for physiological noise by regressing out: (A) cardiac signal, (B) respiratory signal, (C) both cardiac and respiratory, similar to RETROICOR [4], (D) a time series, taken from a mask of large blood vessels within the field of view, (E) manually selected ICA-components, derived from FSL MELODIC ICA. These regressions were performed using the fsl_regfilt program (FSL 5.0). The efficacy of these methods was assessed by generating cross-cortex, depth-dependent profiles of statistical significance, quantified by a z-statistic generated by the FSL FEAT analysis tool. Each profile was compared to the original significance of activation, without any physiological remediation.

Results

Fig. 1A depicts maps of desired tapping induced activity without any removal of physiological noise. In specific ICA components (Fig. 1B) and in maps of cardiac-induced signal changes (Fig. D-E) clear, locally confined structures can be identified (e.g., red arrows). Fig. 2 demonstrates the effect of regressing out the physiological noise as a function of cortical depth. In Fig. 2A, the variance explained from physiological noise is highly dependent on the position within the cortical depth. While respiration induced noise (orange in Fig. 2A) is dominant in CSF areas only, cardiac induced noise (blue in Fig. 2A) additionally affects middle cortical layers. Please note the scaling of the y-axis; cardiac and respiratory noise can only explain less than 1% of the variance.

Discussion

The fact that the physiological regressors explain a limited portion of the overall variance is consistent with previous high-resolution studies investigating physiological fluctuations [2]. These observations are consistent with the interpretation that physiological noise is not the limiting factor at high resolution. It is possible that images are mainly dominated by thermal noise because of the smaller voxel size. This thermal sensitivity may be more significant in resting state analyses, as compared to task-oriented experiments. This non-physiological dominance may also be coming from the fact that the local structures of high physiological noise (shown in Fig. 1) are far from the ROI of the hand knob. Subjective, image based physiological noise cleaning can be difficult to implement effectively, as one may inadvertently remove task-related activity, along with noise. Hence, special care must be taken to ensure that the regressors used here are orthogonal to the tapping induced activity. Other factors that may be influencing these results include our lack of spatial smoothing, as when this is done, the randomly distributed thermal noise averages out, and physiological noise is left dominant.

Conclusion

Our results suggest that in the thermal noise dominated regime of sub-millimeter voxels, depth-dependent physiological noise has a limited effect on fMRI signals. It is entirely possible that thermal variations may be the dominant source of noise, and under this assumption, it is essential to mediate it for effective high-resolution imaging.

Acknowledgements

This research is supported by the NIH Intramural Research program, specifically the Section on Functional Imaging Methods.