1339252 - Systematic Evaluation of Deep Learned Reconstruction of Cardiac Cine MR Images on Volumetric Indices and Image Quality

Using consecutive breath-holds (BH) for ventricular volumetry and myocardial mass assessment with cine balanced steady state free precession (bSSFP) imaging remains a limitation. Reducing the duration of and/or the number of breath-holds is beneficial, especially in children and patients with impaired breath-holding capacity.

Methods:

In this HIPPA compliant IRB approved study, breath-held (BH) fully sampled (FS) bSSFP cine MR images, in ventricular short-axis (SAx) orientation from clinical MR examinations of pectus excavatum patients were analyzed retrospectively. All cine acquisitions were performed with retrospective cardiac gating and prospective arrhythmia rejection. The imaging parameters were:  repetition time/echo time, 2.5–2.7/1.25–1.35 (ms); flip angle, 60°; acquired voxel size, 1.6–1.8 × 1.6–1.8 × 8mm³ (2mm gap); acquired and reconstructed temporal resolution, 22–42 ms (24 phases). FS acquisitions were retrospectively undersampled in the spatial dimension to compressed sensitivity encoding (C-SENSE) acceleration factors (R) of 2 to 8 and reconstructed with the adaptive intelligence based deep learned (DL) prototype algorithm provided by MR vendor [1,2].  Quantitative measures of quality imperceptibility (QI) such as structural similarity index measure (SSIM), multi-scale SSIM (MSSIM), and Peak signal-to-noise ratio (PSNR) were computed between the two consecutive end-diastolic (ED) images of the FS data and between the corresponding images from the FS data the DL reconstructions of the undersampled data for each R and each cardiac phase. One cardiologist and two radiologists scored all images for image quality (IQ) on a five-point scale. For volumetry, DL-based automated biventricular segmentation of all cine SAx images was performed.  Comparative statistical analysis was performed for QI, IQ, biventricular volumes, and left ventricular (LV) mass.

Results: Fifteen patients (12 m / 3 f; 16.8 ± 5.4 years, body surface area, 1.6 ± 0.2 m², heart rate 75.4 ± 16.8 bpm) underwent fully sampled cine bSSFP acquisitions. The SSIM maps for representative SAx images are depicted in Fig. 1. The QI values and edge definition score decreased significantly (p < 0.00001) with increasing R.  For R upto 3, the DL reconstruction QI values were comparable to the corresponding values for the two consecutive ED phase images of the FS data (Fig. 2) and the IQ was scored diagnostically adequate for up to R=5 (Fig. 3). The effect of R was non-significant (p > 0.05, mean difference < 3%) for all biventricular volumetric indices for R< =6. There was significant (p < 0.05) underestimation of LV mass for R >5 (Fig. 3). There was significant (p < 0.02) linear increase in inter quartile range of differences in RV EDV, ESV, SV and EF with increasing acceleration factor.

Conclusion:

Biventricular volumetric indices and LV mass using C-SENSE R = 5 with DL reconstructions were comparable to SENSE R = 2 and image quality was diagnostic for nominally identical spatial and temporal resolution.