Automated Measurement of Native T1 and Extracellular Volume Fraction in Cardiac Magnetic Resonance Imaging Using a Commercially Available Deep Learning Algorithm

Figures

• 

  Fig. 1. Study flowchart. From 755 patients who underwent cardiac magnetic resonance imaging patients with left ventricular hypertrophy and dilated cardiomyopathy were consecutively selected. The exclusion criteria were inadequate image quality or uncertain diagnosis. Healthy volunteers were included from a prospective study. Finally, 36 patients with left ventricular hypertrophy, 32 patients with dilated cardiomyopathy, and 27 healthy volunteers were included.

• 

  Fig. 2. Illustration of the fully automated myocardium analysis. The DL architecture, 2D U-Net, used for myocardium segmentation is shown in the left image. Developed DL models segment the myocardium in native T1 and post-T1 maps. Based on DL segmentation results, RVIPs and reference axes are generated for AHA 16 segments analysis. In the middle image, automatically defined myocardium contours and color-coded reference lines are shown. T1 maps and ECV fractions are converted to the AHA 16 segments model. Automated measurements are summarized in the form of a bull’s eye map as shown in the right image. AHA = American Heart Association, DL = deep learning, ECV = extracellular volume, RVIP = right ventricular insertion point

• 

  Fig. 3. Scatter plots and Bland-Altman plots of automated and reference native T1 values. Automated T1 showed strong correlation with reference T1 in per-subject (r = 0.967 [95% CI 0.951–0.978], slope = 0.934), per-slice (r = 0.980 [95% CI 0.975–0.984], slope = 0.996), and per-segment analyses (r = 0.951 [95% CI 0.946–0.956], slope = 0.965). Automated and reference T1 values were in good agreement in per-subject (bias 9.5 msec, 95% LoA -23.6–42.6 msec), per-slice (bias 10.6 msec, 95% LoA -32.8–53.9 msec), and per-segment analyses (bias 9.6 msec, 95% LoA -61.4–80.6 msec). CI = confidence interval, LoA = limits of agreement

• 

  Fig. 4. Scatter plots and Bland-Altman plots of automated and reference ECV. Automated ECV showed strong correlation with reference ECV in per-subject (r = 0.987 [95% CI 0.980–0.991], slope = 0.943), per-slice (r = 0.968 [95% CI 0.960–0.975], slope = 0.943), and per-segment analyses (r = 0.956 [95% CI 0.951–0.960], slope = 0.935). Automated and reference ECV were in good agreement in per-subject (bias 0.7%, 95% LoA -2.8%–4.2%), per-slice (bias 0.9%, 95% LoA -4.4%–6.2%), and per-segment analyses (bias 0.7%, 95% LoA -5.7%–7.2%). CI = confidence interval, ECV = extracellular volume, LoA = limits of agreement