Predictive signal modeling and multi-rate filtering in accelerated cardiac MRI.

Purpose: True real-time cardiac MRI (CMR), necessary for capturing live cardiac dynamics and imaging irregular cardiac rhythms, remains challenging. In this article, we move toward real-time CMR in multiple reconstruction frameworks via strategies to predict cardiac motion, improve computational efficiency, reduce artifacts, and preserve spatial resolution.

Theory and methods: A published predictive signal model (PMOT) for imaging irregular cardiac dynamics was modified (mPMOT) to enable efficient computation of state-transition matrices for predicting cardiac motion, as training PMOT is computationally expensive. A multi-rate Kalman filter framework was developed to enable computationally efficient reconstructions of high-resolution, large-matrix CMR datasets. Reconstructions were evaluated on multi-coil CMR data in human and swine using multi-rate Kalman filtering and compressed sensing (CS).

Results: Training mPMOT is two orders of magnitude faster than PMOT. Across all datasets and frameworks, mPMOT facilitated high-quality reconstructions of CMR images for different undersampling patterns at acceleration factors of 9 and 13.5. Furthermore, mPMOT substantially reduced temporal blurring artifacts naturally present in CS reconstructions. In swine, mPMOT reduced the mean-squared error of the multi-rate Kalman filter by two orders of magnitude. The multi-rate Kalman filter implementation maintained spatial resolution while reducing computation time from 5439 s to 56 s in select applications.

Conclusion: Our mPMOT is computationally efficient and can be integrated within multiple established reconstruction frameworks to ensure robust tracking and reconstruction for dynamic and real-time CMR applications.