TL;DR
This paper introduces PROM (Phase Recovery from Multiple Wrapped Measurements), a novel algorithm for more accurate and robust velocity estimation in Phase Contrast MRI (PC-MRI). PROM significantly improves upon existing methods by jointly processing all available phase differences, leading to a wider unambiguous velocity range and reduced estimation errors, which is crucial for diagnosing various cardiovascular conditions.
What Problem Does This Paper Address?
Phase Contrast MRI (PC-MRI) is a vital tool for visualizing blood flow dynamics. It encodes velocity information into the phase of the MRI signal. However, this phase measurement is prone to "wrapping," where velocities exceeding a certain threshold (determined by the Venc setting) appear as lower velocities, leading to aliasing errors. Traditional methods often struggle with this, either by limiting the measurable velocity range or by introducing significant errors. Existing approaches for unwrapping these phase measurements are computationally expensive and don't offer guidance on optimizing the MRI acquisition parameters (like Venc settings) to minimize errors. This paper addresses the need for a faster, more accurate, and acquisition-design-aware method for velocity estimation in PC-MRI, especially in the presence of noise and intra-voxel dephasing.
Methods
The core contribution is the PROM (Phase Recovery from Multiple Wrapped Measurements) algorithm. PROM tackles the problem by treating the phase measurements as a system of noisy congruence equations. Key aspects of PROM include:
- Joint Processing of Pairwise Phase Differences: Instead of processing individual measurements, PROM considers all pairwise phase differences simultaneously. This allows it to leverage correlations between measurements and improve robustness.
- Approximate Maximum Likelihood Estimator (MLE): PROM constructs a set of candidate wrapping integers and then efficiently finds the velocity estimate that maximizes the likelihood function. This avoids computationally expensive grid searches.
- Accommodation of Intra-voxel Dephasing and Noise Correlation: PROM explicitly accounts for signal attenuation due to dephasing within a voxel and the inherent correlations in the noise across different measurements and coils. This is achieved by estimating the phase difference noise covariance directly from the data, without needing prior knowledge of coil sensitivities or dephasing factors.
- Optimized Acquisition Design: A significant innovation is that PROM's probabilistic framework allows for the design of optimal Venc settings. By predicting unwrapping error probabilities and estimating the achievable unambiguous velocity range, researchers can tailor acquisition parameters to specific clinical needs and SNR levels.
- PROM+ Post-processing: An extension, PROM+, incorporates spatial information by leveraging the estimated probabilities of wrapping integers. This further refines the velocity map by considering local correlations, leading to improved accuracy.
Datasets
The paper evaluates PROM using:
- Simulated data: Generated to mimic PC-MRI acquisitions under various dephasing scenarios.
- Phantom data: Acquired from a spinning phantom on a 1.5T scanner.
- In vivo data: Collected from a healthy volunteer on a 3T scanner.
The paper does not introduce new benchmark datasets but rather uses existing experimental setups and simulated scenarios to validate its proposed method against established techniques.
Key Insights and Ideas
- The Power of Pairwise Differences: The paper demonstrates that jointly analyzing all pairwise phase differences, rather than just a few, is key to overcoming the limitations of traditional PC-MRI velocity estimation.
- Unwrapping Errors are Probabilistic: By characterizing the probability of unwrapping errors, PROM enables a more principled approach to acquisition design and post-processing, moving beyond simple heuristics.
- Acquisition Design Matters: PROM provides a framework to design the MRI acquisition (specifically, the Venc settings) to optimize for desired velocity ranges, SNR, and acceptable unwrapping error rates, rather than just analyzing existing data.
- Speed and Accuracy Synergy: PROM achieves a significant speedup (e.g., 500x reduction in search space compared to grid search) while simultaneously improving accuracy, making advanced velocity quantification more feasible in clinical settings.
Key Results
- Reduced RMSE: PROM demonstrated a significant reduction in Root Mean Squared Error (RMSE) compared to conventional methods. For a phantom study, PROM showed a 48.5% decrease in RMSE compared to a "dual-venc" technique. In simulations with intra-voxel dephasing, PROM achieved a 10.5% RMSE reduction over ODV and a 25.1% reduction over SDV.
- Wider Unambiguous Velocity Range: PROM can recover velocities across a range significantly larger than twice the highest Venc, thanks to its ability to jointly unwrap multiple phase measurements.
- Computational Efficiency: PROM offers a substantial speed advantage, with its search space for wrapping integers being significantly pruned compared to exhaustive grid searches.
- Improved Unwrapping: The PROM+ post-processing step, leveraging spatial information, further improved velocity estimation and reduced aliasing errors, particularly in challenging scenarios.
Limitations and Future Work
- Assumptions: PROM relies on certain approximations, such as the assumption that the noise follows a joint Gaussian distribution and that the probability of the true wrapping integer being within a certain range is close to 1. While validated, these assumptions might have limitations in extreme noise conditions.
- Extension to 3D/4D Flow: The current work primarily focuses on single-component velocity estimation. The authors mention that extending PROM to estimate all three velocity components simultaneously, and thus handle multi-variable congruence equations, is a direction for future work.
- Model-based Post-processing: While PROM+ uses a non-parametric local quadratic regression, future work could explore more sophisticated model-based approaches for spatial post-processing to further enhance accuracy.
Summary updated