Prospective Respiratory Motion Correction Using Machine Learning and Pilot Tone (PROMPT) in Cardiac MRI

To implement and evaluate the feasibility of a Pilot Tone (PT)-based prospective gating and tracking technique, which uses a long short-term memory (LSTM) neural network to predict respiratory motion from PT signals.

A subject-specific calibration scan consisting of 100 ECG-triggered single-shot images was performed. Respiratory motion was estimated from the images and PT data were processed to extract the respiratory component. The LSTM model was trained to predict respiratory motion from PT signals. During respiratory-corrected scans, PROMPT-predicted slice-shifting parameters were used to update gating information and slice position for each heartbeat before image acquisition. The method was retrospectively evaluated in 12 healthy volunteers, comparing LSTM with linear and polynomial regression models using normalized root mean square error in decibels (NRMSE_dB) and mean absolute error (MAE). PROMPT was then implemented in late gadolinium enhancement (LGE) and compared with free-breathing retrospective gating in 14 patients. Residual in-plane motion was calculated to assess performance.

The LSTM model achieved an NRMSE_dB of -7.20 dB and an MAE of 1.97 mm between predicted and actual motion, demonstrating significantly higher accuracy than either regression models (p < 0.05). The residual in-plane motion in PROMPT-LGE was significantly lower than in FB-LGE (1.22 ± 0.38 mm vs. 1.35 ± 0.48 mm, p = 0.033).

The proposed respiratory motion correction approach was successfully implemented. The LSTM-based predictive model outperformed linear and polynomial regression models. In LGE imaging, PROMPT significantly reduced in-plane motion and showed potential for limiting through-plane motion compared to the clinical protocol.