Automatic image-based respiratory signal extraction in real-time CMR

Chong Chen, Yingmin Liu, Orlando P. Simonetti, Rizwan AhmadarXiv 2020 · DOI · arXiv

Abstract

Accurate quantification of cardiac function in real-time (RT) cardiac MRI (CMR) requires accounting for respiratory motion. This paper presents a fully automatic method to extract respiratory signals and determine their directionality from free-breathing RT cine images. The method utilizes Principal Component Analysis (PCA) for signal extraction and a novel two-step procedure involving Zeroth-Moment Center (ZMC) curves for sign correction. Evaluated on 11 volunteers and 2 patients, the method demonstrated high correlation (0.91–0.97) with manual references and superior robustness compared to existing techniques. This approach facilitates accurate, breath-hold-free cardiac function assessment.

Introduction

Standard cardiac magnetic resonance (CMR) imaging typically relies on multiple breath-holds to mitigate motion artifacts. However, this requirement is often prohibitive for patients with respiratory distress, heart failure, or pediatric patients. Real-time (RT) CMR has emerged as a solution, allowing for free-breathing acquisitions. While RT-CMR captures the heart's dynamics, it introduces the complexity of respiratory-induced variations in cardiac position and volume. To obtain reliable clinical metrics like stroke volume or ejection fraction, heartbeats must be synchronized to specific respiratory phases (e.g., end-expiration).

Existing image-based methods for respiratory signal extraction often face "sign ambiguity"—the extracted signal may be inverted (peaking at inspiration for one slice and expiration for another). Resolving this usually requires manual intervention or unreliable anatomical assumptions. This study proposes a fully automated pipeline to extract and correctly orient respiratory signals without human input.

Methodology

The proposed pipeline consists of two primary stages: signal extraction and directionality determination.

1. Signal Extraction via PCA

The process begins by applying a low-pass filter to the raw image series to suppress high-frequency cardiac motion while preserving the slower respiratory signal. Principal Component Analysis (PCA) is then performed on the flattened image matrix. In free-breathing RT-CMR, respiration is the dominant source of low-frequency intensity variation. Consequently, the first principal component (the first singular vector) effectively captures the respiratory waveform.

2. Directionality Determination

Because PCA signals have an arbitrary sign, a two-step correction process is implemented:

  • Inter-slice Consistency: To ensure all slices in a multi-slice acquisition are synchronized, the algorithm correlates "eigen images" (the spatial representation of the PCA component) between adjacent slices. This ensures that "up" in the signal represents the same physical phase across the entire heart volume.
  • Global Sign Correction (ZMC): To determine the absolute physical meaning of the signal (inspiration vs. expiration), the authors utilize the Zeroth-Moment Center (ZMC). The ZMC represents the "center of mass" of image intensity along the superior-inferior axis. During inspiration, the diaphragm and heart move inferiorly (downward), shifting the ZMC. By maximizing the correlation between the PCA-extracted signal and the ZMC curve, the algorithm anchors the abstract signal to physical anatomy, identifying the inspiratory phase automatically.

Experimental Results

The method was validated using a 1.5T scanner on 11 healthy volunteers and 2 clinical patients. A 10-slice short-axis stack was acquired for each subject. For comparison, a reference respiratory signal was generated by manually selecting a region-of-interest (ROI) and applying non-rigid image registration.

Performance and Robustness

The automated signal showed exceptional agreement with the manual reference, achieving Pearson correlation coefficients (PCC) between 0.91 and 0.97. Notably, the proposed method successfully determined signal directionality in 100% of the cases. In contrast, a recently proposed state-of-the-art method (Unsupervised Respiratory Extraction, or URE) failed in several instances where the lung-diaphragm interface was not clearly delineated, highlighting the robustness of the ZMC-based approach.

Clinical Utility

The researchers demonstrated the method's practical value by quantifying respiration-induced changes in cardiac output. They found that Right Ventricle (RV) end-diastolic and end-systolic volumes were significantly higher during peak inspiration compared to peak expiration (p < 0.01). This confirms that failing to account for respiratory phase can lead to significant measurement errors and potential misdiagnosis in clinical settings.

Discussion and Conclusion

The primary innovation of this work is the use of the ZMC as a physical anchor for PCA signals. This allows the system to bypass the need for manual ROI selection or complex diaphragm tracking. The computational efficiency is also high, with processing times around 10 seconds for a full dataset, making it suitable for clinical workflows.

While the method is robust, its performance may degrade with significant "bulk motion" (patient shifting) or very short scan durations (less than 5 seconds per slice) where a clear respiratory pattern cannot be established. Future work will involve testing the algorithm on a broader range of pathologies, such as patients with irregular breathing patterns like Cheyne-Stokes respiration.

In conclusion, this fully automatic method provides a reliable way to extract and orient respiratory signals in RT-CMR, enabling precise cardiac function quantification in patients who cannot perform breath-holds.

Summary updated

Abstract

Purpose: To develop a fully automatic method for extraction and directionality determination of respiratory signal in free-breathing, real-time (RT) cardiac MRI. Methods: The respiratory signal is extracted by a principal component analysis method from RT cine images. Then, a two-step procedure is used to determine the directionality (sign) of the respiratory signal. First, the signal polarity of all slices is made consistent with a reference slice. Second, a global sign correction is performed by maximizing the correlation of the respiratory signal with the zeroth-moment center curve. The proposed method is evaluated in multi-slice RT cine from eleven volunteers and two patients. The motion in a manually selected region-of-interest is used as reference. Results: The extracted respiratory signal using the proposed method exhibits high, positive correlation with the reference in all cases and is more robust compared to a recently proposed method. To demonstrate its clinical utility, the method is used to identify heartbeats from inspiratory and expiratory phases and, in turn, to account for respiration-induced changes in the cardiac output. Conclusions: The proposed method enables fully automatic extraction and directionality determinations of respiratory signal from RT cardiac cine images, allowing accurate cardiac function quantification.
BibTeX
@misc{chen2020automatic,
  author = {Chong Chen and Yingmin Liu and Orlando P. Simonetti and Rizwan Ahmad},
  title = {{Automatic image-based respiratory signal extraction in real-time CMR}},
  year = {2020},
  eprint = {2012.02828},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2012.02828},
}
Figure