1347184 - Accelerated cardiac cine study using deep learning reconstruction in comparison with the conventional cine

The cardiac MRI study is time consuming due to repeating image acquisitions under the breathholds. Several techniques have been proposed to shorten exam time, and recently, the deep learning (DL) reconstruction technique has developed^1-9,with undersampling cine technique^10-12. The aim of this study is to evaluate the image quality and the interchangeability in the volumetry between these two techniques.

Methods:

A total of seven people including five patietns was enrolled in this study. Cine image stacks in short axis of the left ventricle (LV) were obtained using the  ASSET Cine and DL Cine on a 3T clinical scanner (Discovery MR750 3.0T, GEHC, USA). Three series of DL Cine with different parameters were obtained, including A) reduction factor (RF) = 12 under free breathing (FB), B) RF=12 during one breathhold (R12), C) RF = 9 dividing the whole SX into two slabs during each breathhold (R9). Two observers with more than 3 years of experience in CMR independently evaluated the image quality using 5-point scale (1=poot, 2=acceptable, 3=fair, 4=good, 5=excellent), in terms of the contrast between the cavity and myocardium, signal homogeneity, blurring, clarity and general imaging quality. The cine image data was transferred to an independent post processing software (Ziostation 2, Tokyo, Japan). Two observers independently drew the outer and inner contours of LV wall, and cardiac function was calculated in values of the LV end-diastolic volume (LVEDV), stroke volume (SV), ejection fraction (EF), and LV mass (Mass). The averaged values derived from two readers were used for the comparison. The relationship between variables was shown using scatter plots with Spearman’s correlation coefficients. Differences between two techniques were also evaluated using Bland-Altman plots, expressing mean difference and 1.96 x standard deviation of the difference.

Results:

The representative images are demonstrated in Figure 1 (Lt). The general imaging quality scores of DL Cine were 4.0 in FB, 4.4 in R12 and 4.6 in R9, which were lower than the score of 4.9 in ASSET cine. All scores were more than 4.1, excluding the blurring score of FB (3.3) (Fig 1 [Rt]). On the scatter plots between volumetric results from ASSET and DL Cines, strong correlations (r > 0.8) were found, excluding Mass in R12 and SV in FB and R12 (Fig 2). Although comparison of values measured on ASSET and DL Cines demonstrated substantial difference from zero in Mass (-8.8 g [-25.9,8.3], -9.1 g [-45.7, 27.5], -13.3 g [-41.1,14.4] on FB, R12 and R9, respectively). There were intermediate differences from zero in SV (-3.9 mL [-20.5,12.7]) and EF (-4.1% [-15.9,7.8]) on FB. Differences in other comparisons were not significant (Fig 3).

Conclusion:

Although substantial difference was found in LV mass measurements, cardiac cine using DL Cine (BH R9) is a promising method for evaluating cardiac function in two breath holds. In addition, free-breathing technique is also acceptable for cine study for patients with insufficient breathhold.