VAST: Vascular Flow Analysis and Segmentation for Intracranial 4D Flow MRI

TL;DR

This work addresses the bottleneck of translating intracranial 4D Flow MRI into routine quantitative metrics by removing manual segmentation and improving velocity reconstruction under noise and phase aliasing. The authors introduce VAST, an unsupervised pipeline that automatically derives vessel masks from complex 4D Flow data and performs continuity-informed phase unwrapping, outlier correction, and low-rank denoising to yield self-consistent velocity fields. Across synthetic, in vitro, and in vivo datasets, VAST delivers sub-voxel geometric accuracy, substantial reductions in velocity error, and improved divergence residuals, all while running in minutes on a standard CPU. This combination of automation, physics-informed constraints, and efficient computation supports broader clinical adoption of intracranial 4D Flow biomarkers and sets the stage for future extensions to multi-label segmentation and explicit biomarker estimation.

Abstract

Four-dimensional (4D) Flow MRI can noninvasively measure cerebrovascular hemodynamics but remains underused clinically because current workflows rely on manual vessel segmentation and yield velocity fields sensitive to noise, artifacts, and phase aliasing. We present VAST (Vascular Flow Analysis and Segmentation), an automated, unsupervised pipeline for intracranial 4D Flow MRI that couples vessel segmentation with physics-informed velocity reconstruction. VAST derives vessel masks directly from complex 4D Flow data by iteratively fusing magnitude- and phase-based background statistics. It then reconstructs velocities via continuity-constrained phase unwrapping, outlier correction, and low-rank denoising to reduce noise and aliasing while promoting mass-consistent flow fields, with processing completing in minutes per case on a standard CPU. We validate VAST on synthetic data from an internal carotid artery aneurysm model across SNR = 2-20 and severe phase wrapping (up to five-fold), on in vitro Poiseuille flow, and on an in vivo internal carotid aneurysm dataset. In synthetic benchmarks, VAST maintains near quarter-voxel surface accuracy and reduces velocity root-mean-square error by up to fourfold under the most degraded conditions. In vitro, it segments the channel within approximately half a voxel of expert annotations and reduces velocity error by 39% (unwrapped) and 77% (aliased). In vivo, VAST closely matches expert time-of-flight masks and lowers divergence residuals by about 30%, indicating a more self-consistent intracranial flow field. By automating processing and enforcing basic flow physics, VAST helps move intracranial 4D Flow MRI toward routine quantitative use in cerebrovascular assessment.

Figures (13)

• Figure 1: VAST segmentation workflow. Magnitude images are denoised via low-rank tensor decomposition to initialize a background mask using adaptive thresholding. At each iteration, background statistics are updated to compute magnitude-based and phase-based (SDM) background likelihoods. These likelihoods are fused into a combined background likelihood map using a spatially regularized weighting field (total-variation regularization) to preserve vessel boundaries while promoting spatial coherence. Adaptive thresholding yields an updated flow mask, which is refined by vessel isolation and hole filling. The procedure repeats until convergence (Appendix \ref{['app:likelihood_fusion']}).
• Figure 2: VAST velocity reconstruction workflow. Within the segmented flow volume, phase images are unwrapped under a mass-continuity constraint to obtain continuity-consistent velocity fields. Localized spurious voxels are corrected using universal outlier detection, and the resulting 4D velocity field is denoised in a low-rank POD basis to exploit temporal coherence. Phase unwrapping, outlier correction, and POD denoising are applied iteratively until the reconstruction stabilizes under an energy-based criterion (Appendix \ref{['app:iteration']}).
• Figure 3: Synthetic ICA aneurysm segmentation. PCD, SDM, and VAST surfaces are overlaid on the ground-truth geometry for (a) two noise levels ($SNR=20$ and $SNR=2$) and (b) two velocity-encoding settings ($v_{\text{enc}}=0.4\,v_{\max}$ and $v_{\text{enc}}=0.2\,v_{\max}$).
• Figure 4: Surface-based segmentation error on the synthetic ICA aneurysm. Mean Euclidean surface distance between each automated segmentation (PCD, SDM, VAST) and the ground-truth mesh across (a) SNR and (b) $v_{\text{enc}}$. Distances are normalized by the smallest voxel dimension.
• Figure 5: Volumetric segmentation metrics on the synthetic aneurysm dataset. Accuracy, F1-score, and Jaccard similarity (definitions in Table \ref{['tab:seg_scores_definition']}) are shown for PCD, SDM, and VAST across (a) noise levels (SNR sweep) and (b) velocity-encoding settings ($v_{\text{enc}}$ sweep).