Introduction

With the growing interest in neurofluid dynamics, researchers are increasingly studying large cerebral arteries to better comprehend hemodynamics and their link to cerebrospinal fluid dynamics. Currently, time-of-flight (TOF) imaging is the gold standard for cerebral artery segmentation but is unavailable in most existing magnetic resonance imaging (MRI) databases. Notably, functional MRI (fMRI) signals in large cerebral arteries exhibit large pulsation-driven fluctuations^1,2,3. We leveraged this behavior and developed a data-driven approach for robust and automatic arterial segmentation. To evaluate its reliability, reproducibility, and robustness across different scanning protocols, we systematically validated the results using a local dataset with TOF images and a public dataset with 430 subjects from the HCP aging cohort.

Methods

Data-driven cerebral artery segmentation:
We applied a two-step approach to segment the large cerebral arteries using fMRI with simultaneous finger photoplethysmography (PPG) as detailed below.

Step 1: Preliminary artery segmentation
In the first step, we aimed to identify which voxels show pulsatile signal patterns. We excluded the initial 10 volumes from the rs-fMRI signal to reduce T1-relaxation effects and then performed detrending and filtering. This involved the application of a Butterworth high-pass filter with a cutoff frequency of 0.005Hz. Subsequently, we randomly split the fMRI signals into two temporal subsets and retrospectively aligned each to the cardiac cycle resulting in a voxel-wise cardiac-aligned signal pair (Figure 1, a-c). Then we computed the voxel-wise correlation coefficient between each pair. This process was repeated 100 times, and the mean voxel-wise correlation was used to create a robust correlation map. A higher correlation coefficient corresponded to a higher likelihood of arterial regions as these regions expressed increased pulsatility compared to signals from the rest of the brain. We selected voxels with a high correlation coefficient using a data-driven cutoff, which was based on three standard deviations from the mean. These voxels were the preliminary artery region (Figure 1, d-e). By taking the mean waveform of the preliminary artery region, we constructed a global mean waveform that presents pulsatile signal dynamics within the arterial region (Figure 1, f).

Step 2: Refined artery segmentation
In the second step, we applied voxel-wise correlation with the global mean waveform. The voxels showing a high correlation coefficient (>0.5) with the global mean were considered major cerebral arteries (Figure 1, g-h).

Data Acquisition:
All scans were collected on 3T Prisma Siemens scanners. Two cohorts were included: 1) Four local subjects (age 21-36) with four repeated resting state fMRI scans and corresponding TOF images, and 2) 430 subjects (age 36-100) from the HCP aging cohort with two anterior-to-posterior and two posterior-to-anterior phase encoding resting state scans. Detailed scanning parameters for two cohorts are listed in Figure 5. Finger PPG was simultaneously recorded throughout the fMRI acquisitions. All PPG signals were quality-checked by an in-house algorithm before being included in the analysis.

Validation and reproducibility:
Three sets of analyses were conducted to evaluate the overall performance of our artery segmentation results. First, we qualitatively compared our segmentation results of the four local subjects with their TOF images to validate their reliability, where TOF was brought to the fMRI space. Second, we conducted an intraclass correlation (ICC) analysis on the volumes of the artery segmentation masks generated for subjects from the HCP aging cohort to evaluate the segmentation reproducibility. Only subjects with all four replicates who passed a PPG quality check are included in the analysis. Finally, comparing the results from both cohorts, we evaluated the sensitivity of our approach to the variations in imaging protocols.

Results

Our data-driven artery segmentation results strongly agreed with TOF images across all four local subjects in the 3D projection view (Figure 2). Among the HCP aging participants, the intraclass correlation coefficient for the volumes of artery segmentation across four repeated scans was 0.79 (95% CI: 0.761 to 0.816, p<0.001), indicating a high level of reproducibility. Finally, when comparing the segmentation results between local data and the HCP aging data scanned with different temporal and spatial resolutions, our method showed good consistency and reproducibility between diverse imaging protocols as illustrated in Figure 3.

Discussion

We developed a robust, automatic, data-driven approach for large cerebral artery segmentation. Our approach revealed a strong agreement with the gold standard (TOF) while demonstrating high reproducibility across MRI spatial and temporal resolutions. Our approach showed its potential as a valuable alternative for cerebral artery segmentations, where TOF imaging data is not accessible. In addition, the adaptability to different MRI resolutions highlighted its broad applicability across various scanning protocols.