1343022 - Deep Learning Imaging Reconstruction to acquire more rapid 2D Cine Cardiac MRI versus traditional SSFP- validation of ventricular function

Cardiac MR is traditionally a labor and time intensive examination that requires long breath hold and acquisition time.  The utility of cardiac imaging continues to increase and is the gold standard for the evaluation of cardiomyopathies, congenital heart diseases, and viability.  Ventricular function/volumetry is the hallmark of cardiac MR and especially time consuming.  In a time when access to cardiac MR is increasingly limited, rapid function imaging is necessary for patient comfort, patients with a limited capacity for breath holding, and patient throughput.  Leverage of deep learning (DL) image reconstruction allows for more rapid cine imaging while maintaining image quality.

Methods:

From February-June 2022, consecutive patients presenting for cardiac MR examinations underwent standard short axis (SA) steady-state free precession (SSFP) acquisition, followed by a matched SA stack utilizing DL Cine (acquisition over 3R-R intervals).  Volumetric measurements were performed independently on both standard SSFP and DL Cine utilizing Arterys software.  Image quality was scored on a 5-point Likert scale.  DL Cine uses an unrolled reconstruction network that consists of 8 iterative blocks including both data consistency update and trainable convolutional neural network [1,2]. DL Cine network is trained on 232 fully sampled 2D FIESTA CINE dataset, with a retrospective variable density k-t sampling scheme [3]. l1 loss between ground truth and DL Cine reconstruction images is used as the loss function for network training.

Results:

63 patients (ages 15-85 years) were evaluated.  Mean and standard deviation between SSFP and DL Cine function demonstrated negligible difference (LVEDV 145.9 +- 0.29, LVESV 69.0 +- 0.03, RVEDV 126.5 +- 0.45, RVESV 60.7 +- 0.68, LVEF 54.2 +- 0.15, RV EF 53.1 +- 0.34).  Image quality of SSFP averaged 3.9, while DL Cine averaged 4.2 (Wilcoxon paired signed rank test p value, 0.001).  Average acquisition time was 8 breath holds (2 slices each) with 20 seconds duration for standard SSFP (2.7 minutes gradient time, 6 minutes total), versus 3 breath holds with 16 seconds duration for DL Cine (total 0.9 minutes gradient time, 1.5 minutes total).

Conclusion:

Utilization of DL image reconstruction to acquire SA cine images (acquisition over 3R-R intervals) demonstrated faster acquisition times with no significant difference in overall ventricular function/volumetry, all while maintaining image quality.  Although our study is of relatively small sample size, the utilization of DL image reconstruction affords the ability to save significantly on breath hold and acquisition time, while maintaining image quality.  The overall scan time for a SA stack was reduced 4-fold with DL Cine. These findings are promising in the need for decreasing overall scan and breath hold time for routine cardiac MR examinations.  Further research will need to be performed to leverage additional acceleration (acquisition over less than 3 R-R intervals), valvular evaluation, and additional imaging planes.