⚡ electrical engineering

Extended Synthetic Validation of Optical Flow-Based Velocity Field Reconstruction for AAA Cine-MRI: Angular Velocity, Noise Robustness, and Dense Flow Comparison

This study extends the validation of Pyramidal Lucas–Kanade optical flow for AAA cine-MRI by introducing a new synthetic phantom with closed-form ground truth to demonstrate that the method's optimal window size is motion-regime dependent, that it outperforms dense Farneback flow in noise robustness, and that it accurately recovers angular velocities matching real-patient data.

Original authors: Hanae Soulami

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hanae Soulami

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to watch a movie of a beating heart, but the camera is a bit blurry, and the heart is moving fast. You want to know exactly how fast the blood vessel walls are stretching and twisting. This paper is about a new way to use computer vision to "track" that movement, specifically for a dangerous condition called an Abdominal Aortic Aortic Aneurysm (AAA), which is a bulging weak spot in the main artery of the belly.

Here is the story of the paper, broken down simply:

The Problem: The "Original Movie" Was a Bit Rough

In a previous study, a researcher named Hanae Soulami tried to use a computer trick called Optical Flow to track the movement of an artery in a real patient's MRI scan. Think of Optical Flow like a game of "Where's Waldo?" for pixels. The computer looks at a picture of the artery at one moment, then looks at the next picture a split-second later, and tries to guess how every little dot (pixel) moved.

The original study worked okay, but it had some holes in its logic:

It only looked at one patient.
It only tested the computer's "guessing game" on a simple spinning disk, not a realistic, wiggly artery.
It didn't check if other, more complex computer methods could do a better job.
It couldn't prove if the computer was actually measuring the artery's twisting (angular velocity) correctly because there was no "answer key" to check against.

The New Experiment: Building a "Perfect" Fake Artery

To fix these holes, the author built a Synthetic AAA Phantom.

The Analogy: Imagine you are a video game designer. Instead of filming a real car crash (which is messy and hard to control), you build a perfect, digital crash in a simulator where you know exactly how fast the car was going and how much it spun.
The Reality: The author created a fake, computer-generated artery that pulses and twists. Because they built it from scratch, they know the exact truth (the "Ground Truth") of how fast it is moving and spinning. They even added "static" (noise) to the image to mimic the grainy quality of real MRI scans.

The Contest: Two Computer Methods Fight

The paper pits two different computer tracking methods against each other using this fake artery and a simple moving disk:

The "Spotter" (Pyramidal Lucas-Kanade / PLK): This method is like a detective who only tracks specific, high-contrast features (like the corners of a building). It's fast and good at ignoring noise, but it doesn't track every single pixel.
The "Painter" (Farneback Dense Flow): This method is like a painter trying to fill in the movement for every single pixel in the image. It gives a complete picture but can get confused if things move too fast or if the image is too grainy.

What They Found (The Results)

1. The "Window Size" Matters
The computer needs to look at a small square of pixels to guess movement. The paper found that the size of this square depends on how the object is moving.

Analogy: If you are watching a race car zoom by, you need a wide net to catch it. If you are watching a snail crawl, a tiny net is better.
Result: For the simple moving disk (fast jumps), a smaller window worked best. For the wiggly artery (slow, continuous pulsing), a larger window was needed to get the best accuracy.

2. The "Spotter" vs. The "Painter"

Big Jumps: When an object moves a lot in one frame (like the moving disk), the "Painter" (Farneback) gets completely lost and fails. The "Spotter" (PLK) handles it perfectly.
Small Movements: When the artery is just gently pulsing, both methods do a decent job, but the "Spotter" is still slightly more accurate and handles the "grainy noise" of the MRI much better.
Noise: If you make the MRI image very grainy (noisy), the "Painter" falls apart quickly. The "Spotter" keeps working, just a little less accurately.

3. The "Reset" Button
The original study used a "Periodic Reset" strategy.

Analogy: Imagine you are walking a dog on a leash. If you don't check your position against a map every few minutes, you might drift off course without realizing it. The "Reset" is like checking the map to make sure you haven't drifted.
Result: On the perfect fake artery, the computer didn't drift much, so the "Reset" didn't seem to change the numbers much. However, the author argues it's still a good idea to use it because it ensures the computer tracks the entire artery and doesn't lose points along the way.

4. The Twist (Angular Velocity)
The most exciting part is that the fake artery was designed to twist. The computer successfully measured this twisting speed.

Result: The computer calculated a twisting speed of about 0.20 to 0.23 radians per second. This matched the range found in the real patient from the original study. This proves the computer isn't just guessing; it's actually measuring the twist correctly.

The Bottom Line

This paper didn't invent a new medical treatment, but it validated the tools.

It confirmed that the "Spotter" method (PLK) is the better choice for tracking arteries in MRI scans, especially when the image is noisy or the movement is tricky.
It proved that the computer can accurately measure how much an artery is twisting, not just stretching.
It provided a "test track" (the synthetic phantom) that other researchers can use to test their own tracking software in the future.

In short: The author built a perfect fake artery to prove that their computer method is a reliable "spotter" for measuring the dangerous movements of a real human artery.

Technical Summary: Extended Synthetic Validation of Optical Flow-Based Velocity Field Reconstruction for AAA Cine-MRI

Problem Statement
The reconstruction of velocity fields from standard abdominal aortic aneurysm (AAA) cine-MRI offers a pathway to hemodynamic risk stratification beyond simple diameter measurements. While Phase-Contrast and 4D flow MRI are reference standards, they require specialized acquisition. Standard cine-MRI, acquired routinely, can theoretically be analyzed using optical flow (OF) at no additional cost. However, a previous study by the same author [9] utilizing Pyramidal Lucas–Kanade (PLK) sparse optical flow on a single AAA patient was limited by a lack of dense-flow comparison, absence of angular velocity validation, and reliance on a single patient dataset. This extension addresses these gaps by rigorously validating PLK against dense optical flow methods and establishing a synthetic ground truth for both linear and angular velocities.

Methodology
The study employs a multi-faceted validation approach combining real-world data reproduction, independent benchmarks, and a novel synthetic phantom:

Data Sources:
- Real Patient Data: Reproduction of results from the original study [9] using short- and long-axis cine-MRI of one AAA patient (Rigshospitalet, 2020).
- Synthetic Disk Benchmark: An independent benchmark involving an irregular-speed translating disk to test sensitivity to window size ($winSize$) and noise.
- Synthetic AAA Phantom: A new, literature-grounded phantom constructed using closed-form ground truth for both linear and angular velocities. The phantom parameters (circumferential wall strain, voxel spacing, lumen diameter, and angular velocity) are derived from a real, open-access 4D-CTA dataset of 10 AAA patients (Jamshidian et al., 2025). The phantom simulates pulsatile dilation and rigid anti-clockwise rotation, with Rician noise added to mimic MRI magnitude images.
Algorithms & Configurations:
- Sparse Optical Flow (PLK): Pyramidal Lucas–Kanade with a 4-level Gaussian pyramid. A sensitivity sweep of $winSize$ (ranging from 4 to 30) was performed.
- Dense Optical Flow (Farnebäck): A dense per-pixel flow method using polynomial expansions and a pyramid structure.
- Strategies: Both "periodic-reset" (Eulerian, re-anchoring points to initial positions) and "no-reset" (Lagrangian) tracking strategies were evaluated.
Metrics:
- Disk Benchmark: Mean relative error (%).
- Phantom Validation: Normalized Root Mean Square Error (NRMSE) and Pearson correlation ( $r$ ) against exact ground truth for both linear and angular velocities.
- Noise Robustness: Performance evaluated across noise levels ( $\sigma \in [0, 0.12]$ ).

Key Contributions

Regime-Dependent Optimization: The study demonstrates that the optimal $winSize$ for PLK is not universal but regime-dependent. A smaller window ( $\approx 11\text{--}12$ ) is optimal for large, discrete single-step displacements (disk benchmark), while a larger window ( $\approx 20$ ) is required for small, continuous pulsatile motion (AAA phantom) to average out noise.
Synthetic Angular Velocity Ground Truth: The introduction of a synthetic AAA phantom with exact closed-form ground truth enables the first quantitative validation of angular velocity recovery in this context.
Sparse vs. Dense Comparison: A systematic comparison reveals that PLK and Farnebäck are complementary rather than interchangeable. PLK is superior for large isolated displacements and high-noise environments, whereas Farnebäck provides full-frame coverage but fails on large single-feature displacements against static backgrounds and is more noise-sensitive in small-motion regimes.
Validation of Physiological Consistency: The recovered angular velocity peak from the phantom ( $\approx 0.20\text{--}0.23$ rad/s) aligns with the range observed in the original real-patient study ( $+0.13$ to $+0.22$ rad/s) and existing 4D-flow literature.

Results

Disk Benchmark: PLK achieved a minimum error of $<0.5\%$ at $winSize \approx 11\text{--}12$ , reproducing the original study's findings. Farnebäck failed uniformly ( $\approx 83\text{--}85\%$ error) due to violations of the local smoothness prior when an object displaces more than its diameter per frame.
Linear Velocity (Phantom): Both methods captured the systolic peak correctly. PLK outperformed Farnebäck with lower NRMSE ( $28.4\%$ vs. $32.4\%$ ) and higher correlation ( $r=0.93$ vs. $0.91$).
Angular Velocity (Phantom): PLK achieved an NRMSE of $15.4\%$ ( $r=0.91$ ), significantly outperforming Farnebäck ( $25.0\%$ NRMSE, $r=0.71$ ). The recovered peak velocities matched the physiological range of the original study.
Reset Strategy: On the synthetic phantom, the "reset" and "no-reset" strategies performed nearly identically. The authors note this is a non-confirmatory finding regarding drift magnitude, as the phantom's motion scale did not reproduce the drift accumulation seen in real data. However, the reset strategy is still recommended for ensuring full-lumen coverage and preventing void formation.
Noise Robustness: PLK demonstrated superior robustness to noise compared to Farnebäck. At high noise levels ( $\sigma=0.12$ ), PLK's linear velocity NRMSE was $94\%$ versus $162\%$ for Farnebäck.

Significance and Claims
The paper claims to provide a concrete, actionable path forward for optical flow-based velocity reconstruction in AAA analysis. By validating the original study's findings against a synthetic ground truth and comparing them with dense flow methods, the authors establish that:

PLK is the robust choice for quantitative tracking in cine-MRI, particularly where large displacements or noise are present, provided the window size is tuned to the specific motion regime.
Dense flow (Farnebäck) is suitable for qualitative full-field visualization but is not a drop-in replacement for sparse tracking in this specific application.
The periodic-reset strategy remains valid for maintaining coverage, even if drift magnitude was not fully reproduced in the synthetic model.
Future work is anchored to a specific, open-access multi-patient dataset (Jamshidian et al., 2025), turning the "single-patient" limitation of the original study into a defined next step. The authors explicitly state that multi-patient validation, 4D flow integration, and learning-based optical flow remain open but are now grounded in identified resources.