ADAPTIVE RR PREDICTION FOR CARDIAC MRI

Julien Oster, Olivier Pietquin, Gilles Bosser, Jacques Felblinger, Adaptive Rr Prediction2015

TL;DR

This paper tackles a critical challenge in cardiac MRI: the unpredictable changes in heart rate during breath-holds, which degrade image quality. They propose a simple, clinically applicable method using a Kalman filter to predict upcoming heartbeats by combining respiratory signals and ECG R-wave amplitude, leading to more accurate MRI parameter adjustments and potentially better image quality.

What Problem Does This Paper Address?

Cardiac MRI relies on precise timing to capture images of the heart, which is constantly moving. Traditionally, MRI sequences are synchronized to the patient's heartbeat, specifically the R-wave of the ECG. This synchronization requires setting acquisition parameters (like trigger delays and inversion times) before scanning begins, assuming the heart rate will remain constant. However, patients are often asked to hold their breath to minimize motion artifacts. During a breath-hold, a patient's heart rate naturally fluctuates significantly, invalidating the initial assumptions and leading to suboptimal image acquisition, especially for techniques like black-blood imaging. This can result in ghosting artifacts and even make certain imaging sequences impossible for patients with higher heart rates. The paper aims to solve this by predicting heart rate variations in real-time to adapt MRI parameters dynamically.

Methods

The core of the proposed solution is an Adaptive RR Prediction method. Instead of assuming a constant heart rate, they model the RR interval (the time between consecutive heartbeats) as a dynamic process. They employ a Physiological Multi-Input Regression (PMIR) model, which is essentially a linear combination of various physiological signals.

Specifically, they use:

  • Previous RR intervals: The historical heart rate data.
  • Respiratory signals: Measured using a pneumatic belt on the abdomen.
  • R-wave magnitude: The amplitude of the R-wave in the ECG, which is known to correlate with respiration.

These inputs are combined to predict the next RR interval. To estimate the parameters of this model in real-time, they utilize a Kalman filter. The Kalman filter is well-suited for this task as it can recursively estimate the state of a system (in this case, the heart rate dynamics) from a series of noisy measurements. The paper also mentions that the first breath-hold is used to initialize the model parameters. They tested three variations of the PMIR model, differing in the number of input signals and their orders.

Datasets

The study was evaluated on data collected from five healthy subjects in a clinical MRI setup. Each subject performed two breath-holds in the inspiration position and two in the expiration position. The data collected included ECG signals (for R-wave detection and magnitude), respiratory belt signals, and MRI acquisition timing information. No new benchmark datasets were introduced; the evaluation was performed on data acquired specifically for this study.

Key Insights and Ideas

  • Breath-holds are not heart-rate-stable: The assumption of a constant heart rate during breath-holds is demonstrably false, significantly impacting MRI acquisition timing.
  • Heart rate variation is subject- and mode-dependent: The paper found that heart rate variations during breath-holds are not reproducible between inspiration and expiration modes, and also vary between subjects. However, for a given subject and breath-hold mode, the pattern of heart rate change tends to be similar.
  • Predictive modeling is feasible and beneficial: By modeling heart rate variations using physiological signals, it's possible to predict upcoming RR intervals with reasonable accuracy.
  • Kalman filtering is a practical approach: The Kalman filter provides an efficient way to estimate the model parameters online, making the method suitable for real-time clinical application.
  • Combining respiratory and ECG signals is key: The inclusion of both respiratory belt data and R-wave amplitude provides richer information for predicting heart rate changes, especially during free breathing.

Key Results

The primary result is the demonstration that their adaptive prediction method significantly improves the accuracy of RR interval prediction compared to a constant heart rate assumption. They quantified this improvement using the Mean Squared Error (MSE). The normalized MSE results showed that their PMIR models achieved improvements that were approximately six to seven times better than using a constant RR estimate.

  • PMIR 3 (using more parameters) performed best for inspiration breath-holds.
  • PMIR 2 (using fewer parameters but incorporating R-wave magnitude) performed better for expiration breath-holds, likely because the respiratory belt signal is more stable in this mode.
  • The paper also presented example images (Figure 3) that are described as "unfeasible with the clinical applications" using traditional methods, suggesting the potential for improved image quality with their approach.

Limitations and Future Work

  • Model Sensitivity: The authors acknowledge that to adapt quickly with limited data (short breath-holds), the model needs to be very sensitive, which might lead to over-sensitivity in some cases.
  • Sampling of Respiratory Sensor: A significant limitation is the non-uniform sampling of the respiratory sensor, leading to a loss of information. They suggest that using asynchronous methods could improve prediction.
  • Breath-hold quality: The method cannot compensate for imperfect breath-holds or patient movement, which can still lead to motion artifacts.
  • Clinical comparison: Directly comparing image quality between their method and standard methods is difficult because the quality of the subject's breath-hold itself cannot be controlled.
  • Future Work: The authors suggest that complicating the model could increase prediction quality but might also increase computation time beyond clinical feasibility. They also propose using more breath-holds for better parameter estimation and exploring asynchronous sampling methods for the respiratory sensor. They believe their method would be particularly beneficial for free-breathing or double-gating sequences.

Summary updated

Abstract

Cardiac magnetic resonance imaging (MRI) is very challeng-ing due to the perpetual heart movements. This movement is pseudo-periodic and implies several issues for image acqui-sition. Building a single image requires several shots, done on a specific timing of the cardiac cycle. Nowadays the heart rate is estimated before imaging and then it is assumed not to evolve during acquisition. Additionally, in order to remove motion artifacts, the patients are asked to perform breath-holds. Unfortunately, while performing a breath-hold, the heart rate is changing in a significant way. To address this problem in the framework of clinical applications, we pro-pose a simple method to predict the out-coming RR interval, in order to compute adapted MR parameters. The RR in-terval is the time separating two consecutive R waves and corresponds to one cardiac cycle. Due to its simplicity this method is clinically applicable. The prediction is performed by modeling the heart rate variation as a linear combination of different sensors (here, respiratory sensor and amplitude of R wave). We evaluated this method on five healthy subjects in a clinical setup.
BibTeX
@misc{oster2015adaptive,
  author = {Julien Oster and Olivier Pietquin and Gilles Bosser and Jacques Felblinger and Julien Oster and Olivier Pietquin and Gilles Bosser and Jacques Felblinger and Adaptive Rr Prediction and Julien Oster and Olivier Pietquin and Gilles Bosser and Jacques Felblinger},
  title = {{ADAPTIVE RR PREDICTION FOR CARDIAC MRI}},
  year = {2015},
}
Figure