Introduction

The respiratory navigator echo technique is widely used for free-breathing cardiac and abdominal MRI with minimal respiratory motion artifacts. Navigator echo enables direct respiratory motion detection of the diaphragm and can be used for more accurate respiratory triggering/gating than other techniques with extra devices such as bellows. However, respiratory motion detection with navigator can be degraded due to various causes (e.g. inappropriate navigator tracker positioning), resulting in severe motion artifacts. In this study, we propose a deep learning-based respiratory navigator (DLnav) technique which uses a convolutional neuronal network (CNN) for respiratory motion detection.

Methods

Respiratory Gating with Deep Learning-Based Navigator Echo: The pulse sequence was based on the previously reported navigator-gated 3D SPGR sequence¹ (Figure 1a). As shown in Figure 1, a navigator pulse sequence followed an imaging block for acquisition of pencil-beam navigator echo signals. The multi-channel navigator signal data was then transferred to the real time processing unit. The navigator data was used as an input to the CNN module, which includes 4-layer CNN regression model pre-trained with 1115 navigator data sets. The CNN module output a single diaphragm position value (one of the respiratory waveform points in Figure 1), and the value was passed to the pulse sequence and used for imaging data acceptation/rejection in real time (prospective respiratory gating). The navigator processing was performed in 24 ms including the navigator pulse sequence and diaphragm position value calculation with the CNN module.
Data Acquisition: We performed experiments on a GE 3 T Pioneer imaging system using floating anterior and embedded posterior coil arrays. The navigator-gated 3D SPGR scans were performed with a healthy volunteer using a 20-mm diameter pencil-beam navigator. 3D data were acquired in the axial orientation with imaging parameters: ARC acceleration factor = 2 × 1, TR/TE = 3.2/1.5 ms, FOV = 40 × 40 cm², matrix = 256 × 192, slice thickness = 4.4 mm, receiver bandwidth = ±83.3 kHz, flip angle = 10°. The navigator-gated 3D SPGR scan was performed twice, one with the normal tracker positioning (Figure 2a) and the other with the tracker position intentionally shifted onto the heart (Figure 2b). Non-navigator-gated 3D SPGR scan was also conducted for comparison purpose. Approximate scan time was 40 s and 12 s for navigator-gated and non-gated 3D SPGR, respectively.
Data Analysis: For respiratory waveform evaluation, a signal-to-noise ratio of respiratory-like frequencies² (SNR_R) was measured. 10-s respiratory waveform data were Fourier transformed in the time domain. The maximum value of the power spectrum within the frequency range for respiration (0.1–0.4 Hz) was divided by the mean value for the frequency range of 0.4–3.0 Hz. This evaluation was also performed for the conventional navigator processing method using the same navigator data sets, and the least-squares error method³ was chosen as the conventional one. For image evaluation, sharpness⁴ of the left portal vein was quantitatively calculated. Edge values were calculated for both sides of the vessel using a first-order derivative, and final sharpness was defined as the average score of both sides divided by the center of vessel signal intensity.

Results

Figure 3 shows the waveforms detected with the conventional and DLnav methods and their SNR_R values. The waveforms were well synchronized with the actual respiratory motion and SNR_R was higher than 20 for both methods when the tracker was positioned normally. The conventional method’s waveform was degraded (SNR_R = 2.2) when the tracker was on the heart due to undesired signals in the superior area, but DLnav maintained a good synchronization with the respiratory motion (SNRR = 12.8). Figure 4 shows 3D-SPGR images. DLnav images reduced motion artifacts and improved the sharpness. Quantitative sharpness values of the portal vein shown in Figure 4 were 0.42, 0.30 and 0.18 for DLnav with normal tracker, DLnav with tracker on the heart and non-gating, respectively.

Discussion

We have developed the DLnav technique and have shown that navigator-gated 3D SPGR with DLnav is feasible. The respiratory waveform had a high SNR_R even when the tracker was positioned badly on the heart. Though DLnav images had higher sharpness values for two different tracker potions than the non-gated image, comparison between different tracker positions shows that the tracker on the heart resulted in blurring even with DLnav when compared to the normal tracker position. However, the motion detection and resultant 3D SPGR image quality can be improved further with increased number of training navigator signal data sets of various scan conditions. Clinical evaluation with a large number of subjects is required in future.

Conclusion

The DLnav technique is feasible and improved robustness of free-breathing abdominal MRI in the volunteer scan.

Acknowledgements

No acknowledgement found.