Robust data-driven segmentation of pulsatile cerebral vessels using functional magnetic resonance imaging

Adam M. Wright, Tianyin Xu, Jacob Ingram, John Koo, Yi Zhao, Yunjie Tong, Qiuting Wen

TL;DR

This paper introduces a novel, data-driven method to automatically segment large cerebral arteries and the superior sagittal sinus directly within fMRI data. By leveraging the distinct pulsatile signal patterns of these vessels synchronized with the cardiac cycle, the approach overcomes limitations of traditional methods that struggle with fMRI's inherent distortions, paving the way for more accurate neurofluid dynamics research.

What Problem Does This Paper Address?

Functional Magnetic Resonance Imaging (fMRI) is a powerful tool for understanding brain activity, but its utility in studying neurofluid dynamics and vascular health is hampered by the difficulty of accurately segmenting cerebral blood vessels within fMRI data itself. Existing methods, like Magnetic Resonance Angiography (MRA) or segmentation on structural MRI, often fail because they require separate imaging sequences or struggle with the spatial distortions inherent in fMRI scans. These distortions can misalign vessel structures, leading to inaccurate analyses, especially when trying to understand how blood flow dynamics relate to brain function or waste clearance. This paper aims to bridge this gap by developing a method that can reliably identify and segment these critical vascular structures directly within the fMRI space, eliminating the need for additional scans and avoiding registration errors.

Methods

The core of this work is a data-driven, multi-step segmentation pipeline that exploits the pulsatile nature of blood flow.

  1. Cardiac Alignment: The fMRI signal from individual voxels is randomly split into two temporal subsets and retrospectively aligned to the cardiac cycle, using simultaneous finger plethysmography (PPG) as a reference. A voxel-wise correlation coefficient is calculated between these aligned signal pairs, repeated many times to create a robust measure of pulsatility.
  2. Preliminary Vessel Region Identification: A mean correlation map is generated from the repeated correlations. A participant-specific threshold, set at three standard deviations above the mean correlation across all brain voxels, is used to identify voxels exhibiting high pulsatility. This creates an initial, overly inclusive "preliminary artery region." The average signal waveform from this region is then extracted.
  3. Refined Vessel Segmentation: This step iteratively refines the segmentation. A new correlation map is generated by correlating each voxel's cardiac-aligned fMRI signal with the mean waveform from the preliminary artery region. A data-driven cutoff (based on the 99th quantile of correlations for arteries, and 95th for the SSS) is applied within a general region of interest to identify the final vessel segmentation. This iterative process continues until the segmentation volume stabilizes.

Datasets

  • Local Dataset: A small cohort (n=5, age 21-36) was used for initial validation and comparison with ground truth.
  • Human Connectome Project (HCP)-Aging Dataset: A large dataset (n=422, age 36-90) with four repeated fMRI scans per participant was used to assess the method's robustness and reproducibility.

Key Insights and Ideas

  • Leveraging Intrinsic Pulsatility: The key innovation is using the inherent cardiac-induced pulsatility of blood flow as a direct signal for segmentation, rather than relying on anatomical templates or structural information that can be distorted in fMRI.
  • fMRI-Native Segmentation: Performing segmentation directly within the fMRI space eliminates registration errors that plague methods relying on separate MRA scans, which is crucial for accurate functional analysis.
  • Data-Driven Thresholding: The use of participant-specific correlation thresholds makes the method adaptable to variations in signal quality and individual physiology.
  • Reproducibility Across Age and Scans: Demonstrating high reproducibility across multiple scans and a wide age range in a large dataset is critical for clinical and research applications.

Key Results

  • Qualitative Agreement: The data-driven segmentations showed close qualitative agreement with ground truth (TOF for arteries, manual for SSS) in local participants.
  • Reproducibility: The method demonstrated high reproducibility, with intraclass correlation coefficients (ICC) above 0.7 for both artery (ICC = 0.751) and SSS (ICC = 0.725) segmentation volumes across repeated scans.
  • Quantitative Comparison (Dice Score): While qualitatively good, the Dice scores for artery (0.11 ± 0.03) and SSS (0.28 ± 0.08) segmentation were low. This is attributed to fMRI distortions causing misregistration and the data-driven method potentially including more voxels due to signal sensitivity to hemodynamics.
  • Volume Differences: Artery segmentation volumes were approximately twice that of TOF, and SSS volumes were comparable. Differences were observed between phase-encoding directions, highlighting the impact of fMRI distortions.
  • Spectral Analysis: Confirmed that the segmented voxels indeed exhibited high cardiac pulsatility, validating the underlying principle of the method.

Limitations and Future Work

  • Reliance on PPG: The current method requires simultaneous finger plethysmography. Future work could explore generating PPG signals from fMRI data itself or developing methods that don't require PPG.
  • Pathology Performance: The method's performance in vascular pathologies like stroke remains to be tested.
  • TR Dependence: While tested with TRs ≤ 800 ms, its performance with longer TRs (common in some fMRI protocols) needs further investigation, though cardiac-aligned signals are suggested to be TR-independent.
  • Partial Volume Effects: Due to fMRI's larger voxel size, segmentations may still contain partial volumes of CSF, grey matter, and white matter.
  • Dice Score Discrepancy: The low Dice scores, despite good qualitative agreement, warrant further investigation into the interplay of fMRI distortions, partial volume effects, and the sensitivity of the data-driven approach to pulsatile signals.

Summary updated

Abstract

Functional magnetic resonance imaging (fMRI) captures rich physiological and neuronal information, offering insight into neurofluid dynamics, vascular health and waste clearance. Accurate cerebral vessel segmentation could greatly facilitate fluid dynamics research in fMRI. However, existing vessel identification methods, such as magnetic resonance angiography or deep-learning-based segmentation on structural MRI, cannot reliably locate cerebral vessels in fMRI space due to misregistration from inherent fMRI distortions. To address this challenge, we developed a data-driven, automatic segmentation of cerebral vessels directly within fMRI space. This approach identified large cerebral arteries and the superior sagittal sinus (SSS) by leveraging these vessels' distinct pulsatile signal patterns during the cardiac cycle. The method was validated in a local dataset by comparing it to ground truth cerebral artery and SSS segmentations. Using the Human Connectome Project (HCP) ageing dataset, the method's reproducibility was tested on 422 participants aged 36-90, each with four repeated fMRI scans. The method demonstrated high reproducibility, with an intraclass correlation coefficient > 0.7 in both cerebral artery and SSS segmentation volumes. This study demonstrates that large cerebral arteries and SSS can be reproducibly and automatically segmented in fMRI datasets, facilitating reliable fluid dynamics investigation in these regions.
BibTeX
@inproceedings{wright2024robust,
  author = {Adam M. Wright and Tianyin Xu and Jacob Ingram and John Koo and Yi Zhao and Yunjie Tong and Qiuting Wen},
  title = {{Robust data-driven segmentation of pulsatile cerebral vessels using functional magnetic resonance imaging}},
  booktitle = {Interface},
  year = {2024},
  doi = {10.1098/rsfs.2024.0024},
}

Loading related content...

Figure