Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session C

Poster F9 - K-Means Clustering Shows Spatial Heterogeneity and Discrete Spatial Regions of Scalp Blood Flow When Detecting Cortical Activity using fNIRS

Friday, October 25, 2024

10:00 AM - 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Nabil Othman, Mr

Undergraduate Researcher
Marquette University
Frankfort, Illinois, United States

Co-Author(s)

AZ

Anthony Zinos

Graduate Student
Marquette University and Medical College of Wisconsin, United States
JW

Julie Wagner

Graduate Student
Marquette University and Medical College of Wisconsin, United States
Janine Taitt, Medical Doctor

Clinical Research Fellow
Medical College of Wisconsin, United States

Primary Investigator(s)

SB

Scott Beardsley

Director of Undergraduate Studies, Assistant Professor
Marquette University and Medical College of Wisconsin, United States

Introduction: Subtrack: Optics

Functional near-infrared spectroscopy (fNIRS) images brain function by detecting changes in oxy- (HbO) and deoxyhemoglobin (HbR) in superficial cerebral blood vessels. Relative changes in HbO and HbR are correlated to brain activity via the hemodynamic response. fNIRS measurements of brain activity are contaminated with noise from the scalp vasculature, as light must pass the scalp (and skull) before it passes through the brain. Scalp vasculature adds unwanted physiologic noise, such as high amplitude Mayer wave fluctuations, which can confound measurements of brain activity. If current analysis techniques are unable to parse scalp effects from brain signals of interest, the results become difficult to interpret. While typical fNIRS channels (long channels, 3cm interoptode distance) detect scalp and cortical signals, short separation channels (SSC) are spaced with an interoptode distance of less than 1cm to measure changes in changes in HbO and HbR only from the scalp. These are powerful tools to remove systemic physiological noise from cortical signals. Scalp blood flow signals are often assumed to be spatially homogenous such that only one SSC is needed. However, recent literature suggests this assumption may not be true, as SSC correlation with long channels decreases significantly with distance on the head. This study aims to improve the current fNIRS analysis techniques by determining the relative strength of physiologic signals measured with SSC and how those signals vary with space, to determine how many SSCs are needed to optimize the removal of scalp blood flow from fNIRS measurements of brain activity.

Materials and

Methods: Eight participants performed a five-minute resting state and bilateral finger tapping task (10 blocks of 20-second tapping followed by 20-second rest per hand) with simultaneous fNIRS sampled at 3.125Hz. Whole head sampling of scalp blood flow was performed using 32 SSCs positioned using a 10-20 layout. Motion artifacts are corrected using temporal derivate distribution repair. Spatial heterogeneity was examined by implementing bandpass filtering (4th order, zero-phase, Butterworth) to isolate physiological signals: Ultra Low (ULF, < 0.01 Hz), Very Low (VLF, 0.01-0.05 Hz), Mayer Wave (0.07-0.15 Hz), Respiratory (0.2-0.4 Hz), and Cardiac (0.8-1.5625 Hz). Paired channel-wise cross-correlations were performed within each frequency band and across task paradigms and peak correlations were examined.

K-means clustering was performed to cluster each participant’s 32 SSC based on the frequency-dependent phase angles across paradigms. For each SSC, phase angle in the Fourier frequency domain was calculated using the Fast Fourier Transform (FFT). The optimal number of clusters across SSC phase profiles was determined to be 3 using the elbow method, which identifies the transition to minimal improvement of within-cluster sums of point-to-centroid distances. Each clustering used k-means++ to select initial centroids and square Euclidean distance to determine the best cluster of 1000 replicates.

The relative strength of physiologic components contributing to the scalp signal measured with the SSCs was characterized using the FFT power spectrum. Average power was calculated within each frequency band and subsequently averaged across SSCs to characterize the relative contributions of signals across frequency bands.

Results, Conclusions, and Discussions: K-means clustering revealed spatial structure highlighting hemispheric differences (left or right) and occipital regions (figure 1), suggesting meaningful differences in SSC signals. On average, the left hemisphere cluster contained 11±2 channels, the right hemisphere cluster had 12±2 channels, and the occipital cluster had 10±2 channels.

A frequency-specific pattern of cross-correlations was observed between scalp signals (figure 2) and spatial patterns varied with frequency band. The VLF, Mayer wave, and cardiac bands showed high peak correlation (r >0.7) among frontal channels for bilateral motor SSCs, though this dissipated with distance. Correlations were generally lower (r < 0.3) occipitally and for the respiratory frequency band. The ULF band showed spatially scattered correlations with high correlations (r > 0.7) between channels.

The results suggest scalp blood flow signals are spatial heterogeneous (figure 2), with channels grouped into three discrete regions (figure 1). Cross-correlations decreased with distance between SSCs, suggesting that SSC signals are not globally homogeneous. This may be due to differences in vascular supply; temporal lobes are supplied by the internal carotid arteries and the occipital lobe by the posterior carotid artery. K-means clustering showed spatial differences in SSC phase angles, potentially reducing the accuracy of post-processing techniques. The analyses suggest that fNIRS montages should include at least 3 SSCs in these regions to capture the range of scalp blood flow signals measured by long channels.

Across subjects, the cardiac and ULF frequency bands had the highest power (figure 3). The cardiac band accounted for almost 40% of the power while containing ~49% of the frequency range. The ULF band comprises 0.65% of the frequency range yet contributed approximately 17% of the total power, indicating disproportionate signal magnitude from the ULF band. This demonstrates that low-frequency noise is abundant in fNIRS SSC signals (figure 3). Hemodynamic signals at this frequency are likely due to vasomotion, suggesting that SSC should be high-pass filtered to improve accuracy.

Future studies should explore how varying the number of SSCs within the left, right, and occipital regions impacts GLM analysis. SSC should be high-pass filtered to eliminate low-frequency noise, which accounts for nearly 20% of SSC signals.

Acknowledgements (Optional):