Dynamic Analysis of Cardiac Flow Using 4D Flow MRI (RF_TH_231)

The heart's complex geometry, consisting of large chambers and intricate structures, is powered by strong muscle contractions that create intricate blood flow patterns across various spatial and temporal scales. 4D Flow MRI is effective for analyzing these flows in vivo, however, its utility is limited by resolution, noise, signal averaging, and motion artifacts. Currently, no method effectively addresses all these issues in an automated and physically consistent way.

these issues in an interpretable manner.

Methods:

We developed an automated 4D Flow MRI workflow (Figure 1) for cardiac chamber segmentation, velocity field reconstruction, and flow analysis. Our method, Dynamic Combined Likelihood Iterative Segmentation (dyn-CLIS), leverages MRI signal magnitude and phase data in a maximum likelihood formulation to determine if a voxel is inside or outside of the flow [1][2]. The segmentation is refined iteratively at each cardiac phase using a dynamic thresholding process.

Velocity field reconstruction applies divergence-free constrained least squares optimization, Universal Outlier Detection, and Proper Orthogonal Decomposition [3][4][5], reducing noise and correcting discontinuities. The reconstructed flow is evaluated in terms of physiologically relevant hydrodynamic quantities, such as pressure and wall shear stress. To ensure robust field evaluation via least squares optimization, uncertainty-based weighting is applied [6][7]. Moreover, pathlines are analyzed using two methods: traditional Radial Basis Functions (RBFs) with Runge-Kutta 4-5 interpolation and Bayesian interpolation through approximate Gaussian Process (GP). Flow structures are delineated using three-dimensional Finite Time Lyapunov Exponents (3D FTLE).

Results:

The dyn-CLIS method achieved close alignment with expert segmentations coregistered using fast RAndom SAmple Consensus (RANSAC) and Iterative Closest Point (Figure 1) using a Python library Open3D [8]. Noise near the walls due to partial volume effects and coregistration bias was significantly reduced. A vortex ring in the right ventricle was clearly visualized during the early filling phase, with integrated pressure fields showing a physiological intracardiac pressure difference of approximately 2 mmHg, and a pressure drop at the vortex core. RBF-based pathlines at the end of diastole revealed blood spiraling from the RV to the pulmonary artery due to its curvature.

Conclusion:

Our automated 4D Flow MRI workflow provides accurate and interpretable dynamic segmentation and flow reconstruction, matching expert annotations and reducing velocity field noise. Integrated pressure analysis demonstrated physiological intracardiac pressure differences, and pathline visualization captured coherent flow features. Further improvements using approximate Gaussian Processes and 3D FTLE analysis are planned for future work.