Left ventricular function analysis with deep-learning accelerated cine imaging: A comparison to standard of care segmented cine imaging

Maya Gabbour; Holly Iverson; Kiaran McGee; Phillip Young; Christopher Francois; Alan H. Stolpen; Thomas A. Foley; Tim Leiner; Jeremy D. Collins

doi:10.1016/j.jocmr.2024.101645

Background: Deep learning (DL) sequences are increasingly used in cardiac MRI cine applications. This technique is valuable for assessing ventricular size and function in part to the ability to precisely define the myocardial-blood pool boundary. However, conventional cine sequences require multiple breath-holds, which is time-consuming and challenging, especially for children and patients with arrhythmias, dyspnea, and/or claustrophobia. These factors can lead to artifacts that degrade image quality and hinder accurate cardiac assessment. A recent FDA approved DL-based highly accelerated cine MRI sequence known as Sonic DL (SDL) has been introduced for clinical imaging applications. Our group investigated the three-heartbeat (3RR) MRI SDL method as an alternative to conventional parallel imaging accelerated segmented (CONV) cine imaging. We compared left ventricular ejection fraction (LVEF) results in patients who underwent both SDL and CONV cine cardiac MRI. Methods: In this IRB-approved study, cardiac MRIs performed at 1.5T (Artist or Optima 450w, GE Healthcare, Waukesha, WI) using a multielement cardiac coil were reviewed between April 1 and April 30, 2024. Both SDL (3 seconds / slice) and CONV cine (R=2, 8 seconds / slice) sequences were acquired during the same scanning session. Images were processed on a dedicated platform by a single reviewer blinded to the results from each method (Circle CVi42 v 6.1). LV volumes and estimates of myocardial mass in systole and diastole were compared between methods using linear regression and Bland-Altman analyses. Results: MRIs from 23 patients (age range 25–84 years) were included.  A significant correlation was observed between the reference standard LVEF measured by the CONV and SDL cine methods (r = 0.94, P < 0.001; Figure 1a) with a mean bias of -0.98%, a standard deviation of 4.2%, and limits of agreement ranging (LOA) from -7.3% to 9.26% (Figure 1b). CONV cine yielded higher-quality images with enhanced temporal and spatial resolution compared to SDL cine (Figure 2).  A statistically significant agreement between estimates of myocardial mass in diastole and systole was found (r = 0.97, P < 0.001; Figure 3 a, b). Both CONV and SDL cine demonstrated excellent internal consistency estimating the myocardial mass at systole and diastole (r = 0.99, P < 0.001 for SDL and 0.99, P < 0.001 for CONV cine; Figure 3c, d). Conclusion In a small cohort of adult subjects, the measurement of LVEF was comparable between SDL and CONV cine methods with similar LV mass between phases as a measure of internal consistency. SDL cine could be considered as a convenient, time-saving alternative for dyspneic or claustrophobic patients, significantly reducing table time required for cine imaging. Additional work is ongoing to confirm these results in a larger sample size and to identify factors which may be associated with larger differences between methods.

Left ventricular function analysis with deep-learning accelerated cine imaging: A comparison to standard of care segmented cine imaging

Files and links (1)

Abstract

Details

Metrics