The Impact of Model-based Deep-learning Reconstruction Compared with that of Compressed Sensing–Sensitivity Encoding on the Image Quality and Precision of Cine Cardiac MR in Evaluating Left-ventricular Volume and Strain: A Study on Healthy Volunteers

Figures

• 

  Supplementary Fig. 1. P-values of comparison of visual image quality scores among all reconstructed images. The comparison was performed between the score evaluated by the same observers. Red characters were presented as statistically significant. CS, compressed sensing – sensitivity encoding; DLR, deep learning reconstruction; R, reduction factor.

• 

  Supplementary Fig. 2. Bland–Altman analysis of the left-ventricular stroke volume. Stroke volume did not differ significantly between deep learning reconstruction and compressed sensing – sensitivity encoding, regardless of the reduction factors. CS, compressed sensing – sensitivity encoding; DLR, deep learning reconstruction; GS, gold standard (full-sampling data); R, reduction factor; SD, standard deviation.

• 

  Supplementary Fig. 3. Bland–Altman analysis of the left-ventricular ejection fraction. The ejection fraction did not differ significantly between deep learning reconstruction and compressed sensing – sensitivity encoding, regardless of the reduction factors. CS, compressed sensing – sensitivity encoding; DLR, deep learning reconstruction; GS, gold standard (full-sampling data); R, reduction factor; SD, standard deviation.

• 

  Supplementary Fig. 4. Bland–Altman analysis of the global radial strain. DLR significantly underestimated GRS. In addition, CS tended to underestimate GRS with reduction factor of 6 and 8. CS, compressed sensing – sensitivity encoding; DLR, deep learning reconstruction; GS, gold standard (full-sampling data); R, reduction factor; SD, standard deviation.

• 

  Fig. 1. Representative case of each reduction factor with CS and DLR. The improvement in image quality by DLR was visually confirmed. Compared with CS, DLR improved the distortion of the LV shape and irregularity of the LV contour at high reduction factors of R = 6 and 8. CS, compressed sensing–sensitivity encoding; DLR, model-based deep-learning reconstruction; LV, left ventricle; R, reduction factor.

• 

  Fig. 2. Bland–Altman analysis of the left-ventricular EDV. The precision, defined as the standard deviation of the bias, tended to improve with DLR at R = 8 compared with CS. DLR corrected the slight overestimation of EDV by CS at R = 6 and R = 8. CS, compressed sensing–sensitivity encoding; DLR, model-based deep-learning reconstruction; EDV, end-diastolic volume; GS, gold standard (full-sampling data); R, reduction factor; SD, standard deviation.

• 

  Fig. 3. Bland–Altman analysis of the left-ventricular ESV. The precision, defined as the standard deviation of the bias, tended to improve with DLR at R = 8 than with CS. The slight overestimation of ESV by CS was corrected by DLR at R = 8. CS, compressed sensing–sensitivity encoding; DLR, model-based deep-learning reconstruction; ESV, end-systolic volume; GS, gold standard (full-sampling data); R, reduction factor; SD, standard deviation.

• 

  Fig. 4. Bland–Altman analysis of GCS. The precision, defined as the standard deviation of the bias, was improved with DLR at higher reduction factors such as R = 6 and 8 than with CS. Compared with CS, DLR tended to underestimate or significantly underestimate the GCS. CS, compressed sensing–sensitivity encoding; DLR, model-based deep-learning reconstruction; GCS, global circumferential strain; GS, gold standard (full-sampling data); R, reduction factor; SD, standard deviation.