Prospective Respiratory Motion Gating and Tracking using Machine Learning and Pilot Tone in Late Gadolinium Enhancement (RF_FR_405)

Respiratory motion remains a major challenge in cardiovascular magnetic resonance imaging (CMR). The current free-breathing late gadolinium enhancement (LGE) scans utilize a retrospective motion-corrected method (MOCO-LGE) [1], which acquires 8 single-shot frames and discards the 2 with the largest registration errors. The Pilot Tone (PT) is a motion sensor that has been successfully implemented to monitor physiological motion [2, 3]. In this study, we propose a motion compensation method called Prospective Respiratory mOtion compensation using Machine learning and Pilot Tone (PROMPT) and validate it in 2D LGE.

Methods:

Free-breathing ECG-triggered single-shot images were acquired in sagittal and coronal views over 100 frames to calibrate the PT signal to the 3D respiratory motion of the heart. Images were reconstructed using the scanner’s default reconstruction and were sent, along with the PT data, to the Framework for Image Reconstruction Environments (FIRE) [4] research prototype for processing.  Respiratory motion of the heart was measured from images and PT data were compressed into 4 complex virtual coils [5]. A long short-term memory (LSTM) model was trained to predict the respiratory motion from PT data (Figure 1a).

When applied to LGE scans, PT data were collected and sent to FIRE in real-time. Motion prediction was performed using the pre-trained LSTM to infer the slice shift (Figure 1b). Feedback was then sent to the scanner for gating and slice tracking prior to acquiring the single-shot image.

The proposed method was tested in 7 patients (64.1 ± 12.3 years, 4 females) on a 3T system (MAGNETOM Vida, Siemens Healthineers, Forchheim, Germany). The PROMPT-LGE (averaged over 6 single-shot frames) was collected after completion of routine clinical scans, with prospective gate and track using a ±4 mm end-expiratory window. The imaging parameters and views matched those of MOCO-LGE. For both methods, individual frames were co-registered with MOCO before averaging, and average pixel displacement was calculated within left ventricle to measure the residual in-plane motion. Pairwise Student’s t-tests were performed to compare the extent of displacement.

Results:

Figure 2 demonstrates that PROMPT-LGE source images had less through-plane motion, resulting in less artifacts in the averaged image than with MOCO alone. Notably, because PROMPT-LGE images were acquired after MOCO-LGE, the blood pool signal intensity was lower due to contrast agent washout. PROMPT-LGE required less registration in 6 out of 7 patients compared to MOCO-LGE (Figure 3). While the Student’s t-test did not show a significant difference (p = 0.056), this may be due to the small sample size. The MOCO-LGE technique always rejects 2 out of 8 acquired images, while PROMPT-LGE prospectively rejected an average of 3.5 images based on respiratory motion.

Conclusion:

Prospective respiratory correction using PT can reduce the through-plane motion in LGE images. However, additional patient studies are needed to fully demonstrate its advantages over MOCO-LGE.