PINGS-X: Physics-Informed Normalized Gaussian Splatting with Axes Alignment for Efficient Super-Resolution of 4D Flow MRI

AI summary

Overview

Research area: Medical image analysis and physics-informed machine learning, specifically super-resolution of 4D flow MRI (time-resolved three-dimensional blood velocity imaging). The work sits at the intersection of computational fluid dynamics, physics-informed neural networks (PINNs), and 3D Gaussian splatting (3DGS) from computer graphics.

Technical level: Advanced. The paper includes formal convergence proofs (Theorem 1, Corollary 1), stochastic Big-O error bounds, and high-dimensional Gaussian parameterization alongside a clinical imaging pipeline.

Scope in one sentence: The paper introduces PINGS-X, an explicit spatiotemporal Gaussian representation that replaces slow PINN-based super-resolution of 4D flow MRI with a faster, theoretically grounded, physics-informed alternative.

What This Paper Is About

4D flow MRI measures blood flow velocity non-invasively and is used for cardiovascular diagnostics, but high spatiotemporal resolution requires long scan times, creating a trade-off between acquisition speed and accuracy. Existing super-resolution approaches either need large training datasets (deep learning) or must be retrained per patient with slow implicit neural networks (PINNs). The paper's goal is a per-scan super-resolution framework that keeps physics constraints while training orders of magnitude faster than PINNs.

Key Contributions

1. Normalized Gaussian splatting (NGS) with a formal convergence guarantee. The prediction is redefined as a convex combination of Gaussian-associated physical values, normalizing Gaussian weights so the output is always bounded by the values of the influencing Gaussians. The authors prove universal approximation consistency (Theorem 1) with an explicit convergence rate.

2. Axes-aligned Gaussians (NGS-X). Each Gaussian's covariance is constrained to be diagonal, avoiding the need to parameterize complex 4D rotations, and reducing covariance parameters from 10 to 4 per Gaussian. Corollary 1 shows the convergence guarantee is preserved under this constraint.

3. Gaussian merging for stability and scalability. A density-control strategy builds a graph over Gaussians based on cosine similarity (threshold 0.9) of their normalized influence vectors and merges connected components in a single pass, preventing degenerate solutions caused by duplicate Gaussians in the same region.

4. A rigorous testbed using spatially averaged data. Low-resolution inputs are produced by spatial averaging of high-resolution fields rather than point sampling, which the authors describe as a more realistic and challenging setting than prior work.

Main Findings

• Rosenbrock function convergence: With unnormalized weighting (PIGS), relative error stays high and unstable across Gaussian counts (3.38% at 16 Gaussians, 2.45% at 100, 3.14% at 400, 3.53% at 600). PINGS and PINGS-X converge as N grows, reaching 0.04% (PINGS) and 0.08% (PINGS-X) at 600 Gaussians.

• Synthetic 2D CFD accuracy and speed: PINGS-X achieves the lowest relative L2 error on lid-driven (1.13%), Y-shaped (2.62%), and L-shaped (1.15%) datasets. On the Y-shaped dataset it converges in 4.8 minutes, roughly 40 times faster than the next most accurate model, XPINN (sin), at 195.5 minutes and 3.89%.

• Baseline comparison (synthetic CFD): PINN reached 12.20% (lid-driven), 11.54% (Y-shaped), 11.84% (L-shaped); Siren reached 3.20%, 4.34%, 2.49%; PIG reached 2.33%, 7.06%, 1.48% but required 820.1, 292.9, and 297.8 minutes respectively. Nadaraya-Watson ran in under 1 minute with errors of 3.03%, 5.11%, 3.99%.

• Real 4D flow MRI (carotid phantom): For ×8 super-resolution, PINGS-X achieved 8.98% relative error and 0.95 cm/s RMSE in 2.6 hours, versus Siren at 10.63% and 1.13 cm/s in 30.8 hours and PINN at 25.93% and 2.75 cm/s in 30.1 hours. For ×64, PINGS-X achieved 17.59% and 1.87 cm/s in 3.9 minutes, versus Siren at 18.49% and 1.96 cm/s in 299.2 minutes and PINN at 39.70% and 4.21 cm/s in 259.0 minutes.

• Normalization is necessary for convergence: Removing normalization caused non-convergence (marked NC) on all three synthetic CFD datasets.

• Axes alignment is an efficiency win with negligible accuracy cost: Disabling axes alignment raised training time from 21.9 to 70.3 minutes (lid-driven), 4.8 to 23.3 minutes (Y-shaped), and 6.3 to 22.7 (L-shaped column, labeled in hours in Table 4) with errors changing only from 1.13% to 1.01%, 2.62% to 2.27%, and 1.15% to 1.13%.

• Merging improves stability and efficiency: Removing merging caused experiments to terminate early (below 2000 epochs) due to out-of-memory issues, with reported errors of 1.38%, 1.66%, and 1.12%.

• Robustness to initialization: Varying the initial Gaussian count from 200 to 1600 kept lid-driven error between 1.13% and 1.36%, Y-shaped between 2.51% and 2.74%, and L-shaped between 1.12% and 1.17%.

Methodology in Plain English

The authors borrow the explicit representation idea from 3D Gaussian splatting, where a scene is described by a set of Gaussians rather than a neural network. Instead of rendering images, each Gaussian here carries a physical property (velocity, and optionally pressure), a center in space-time, and a spread along each axis. At any query point, the prediction is a weighted blend of nearby Gaussians' physical properties.

The key change is normalization: rather than summing Gaussians directly (which decays to zero far from any Gaussian and forces the model into unstable fixes like oversized Gaussians or constant additions), the weights are normalized so the prediction is a convex combination of the neighboring values. This makes the representation stable and lets new Gaussians act as refinement control points.

Training starts by initializing Gaussians on a grid over the spatiotemporal domain, with properties taken from the nearest low-resolution data point. The loss combines a data fidelity term (masked mean squared error against low-resolution observations, excluding unmeasured pressure) with a physics term enforcing a dimensionless form of the incompressible Navier-Stokes and continuity equations, weighted by λ = 1.0. Because the representation is explicit, PDE derivatives are computed analytically instead of through expensive backpropagation.

Periodically, the model performs adaptive density control: splitting and cloning Gaussians in high-error regions, pruning or merging redundant ones, and merging Gaussians whose influence profiles are highly similar. Boundaries are handled implicitly by sampling training data on boundaries.

Validation uses three steady-state incompressible 2D CFD simulations (lid-driven cavity, L-shaped, Y-shaped channels from OpenFOAM and StarCCM+) mapping 2D coordinates to velocity and pressure, plus a real 4D flow MRI carotid artery phantom dataset acquired at 0.35 mm spatial and 25 ms temporal resolution, with five frames around peak systole taken from 34 frames per cardiac cycle.

Why This Matters

Impact on research: The paper challenges the assumption that physics-informed learning for high-dimensional flow data must rely on slow implicit neural representations. It provides theoretical grounding (convergence rates for normalized Gaussian sums) that prior physics-informed Gaussian work lacked, and demonstrates that explicit representations can transfer from computer graphics to scientific field reconstruction.

Real-world applications:

• Cardiovascular diagnostics, where high-resolution velocity fields help detect and monitor conditions such as stenosis and aneurysms.
• Reducing MRI scan duration by enhancing rapidly acquired low-resolution scans after the fact.
• Hemodynamic assessment of patient-specific vascular geometries, such as the carotid artery phantom used here.
• Scientific computing workflows where physics-constrained field reconstruction is needed from sparse or averaged measurements.

Industry relevance: Per-patient retraining is a practical barrier to deploying physics-informed super-resolution in clinical settings. PINGS-X reduces this from roughly 30 hours to 2.6 hours on the ×8 carotid task and from 259–299 minutes to 3.9 minutes on the ×64 task, which changes the feasibility calculus for imaging vendors, hospital radiology departments, and medical device companies building post-processing pipelines. The code is publicly available.

Future Directions

• In vivo validation. The real-data experiments use a carotid artery phantom that replicates patient vasculature; the paper does not report results on in-vivo patient scans, and the sensitivity of the approach to physiological variability such as motion artifacts is not established in the provided content.
• Extension beyond the evaluated geometries and equations. The paper's supplementary material lists a Burgers' equation experiment, suggesting exploration of other physical systems; broader generalization to turbulent or non-Newtonian flow is not reported.
• Scaling the merging criterion. The current merge uses a fixed cosine similarity threshold of 0.9 on influence vectors and is applied every 100 epochs; whether adaptive thresholds improve robustness in larger domains is an open question.
• Detailed limitations and future work. The supplementary table of contents includes a dedicated "Limitations and future work" section, but its contents are truncated in the provided material and are therefore not summarized here.

Target Audience

Researchers and practitioners in medical image analysis and physics-informed machine learning, particularly those working on 4D flow MRI, super-resolution, or physics-constrained field reconstruction. It is also relevant to computer graphics researchers interested in transferring Gaussian splatting beyond novel view synthesis, and to clinical translation teams evaluating whether per-patient physics-informed reconstruction is computationally practical. A background in neural fields, PDE-constrained optimization, or kernel methods is helpful for the theoretical sections.

Authors’ abstract

4D flow magnetic resonance imaging (MRI) is a reliable, non-invasive approach for estimating blood flow velocities, vital for cardiovascular diagnostics. Unlike conventional MRI focused on anatomical structures, 4D flow MRI requires high spatiotemporal resolution for early detection of critical conditions such as stenosis or aneurysms. However, achieving such resolution typically results in prolonged scan times, creating a trade-off between acquisition speed and prediction accuracy. Recent studies have leveraged physics-informed neural networks (PINNs) for super-resolution of MRI data, but their practical applicability is limited as the prohibitively slow training process must be performed for each patient. To overcome this limitation, we propose PINGS-X, a novel framework modeling high-resolution flow velocities using axes-aligned spatiotemporal Gaussian representations. Inspired by the effectiveness of 3D Gaussian splatting (3DGS) in novel view synthesis, PINGS-X extends this concept through several non-trivial novel innovations: (i) normalized Gaussian splatting with a formal convergence guarantee, (ii) axes-aligned Gaussians that simplify training for high-dimensional data while preserving accuracy and the convergence guarantee, and (iii) a Gaussian merging procedure to prevent degenerate solutions and boost computational efficiency. Experimental results on computational fluid dynamics (CFD) and real 4D flow MRI datasets demonstrate that PINGS-X substantially reduces training time while achieving superior super-resolution accuracy. Our code and datasets are available at https://github.com/SpatialAILab/PINGS-X.