Left Ventricle Segmentation in Cardiac MR Images via an Improved ResUnet

Figures

• 

  Figure 1. The proposed segmentation model with labels Add (addition), Conv (convolution), DW-Conv (depth-wise convolution), BN (batch normalization), and ReLU (rectified linear unit).

• 

  Figure 2. ROI cropping. The image on the left is the original slice, and the white box represents the ROI containing the LV. The image on the right is the corresponding cropped ROIs.

• 

  Figure 3. Data augmentation. The image on the left is the corresponding cropped ROI, and those on the right are augmented ROIs.

• 

  Figure 4. Depth-wise separable convolution.

• 

  Figure 5. Automatic segmentation (solid green line) and ground truth (dotted red line) in the basial slice, midslice, and apical slice for four types of patients, labeled with Basal (basal slice), Mid (midslice), Apical (apical slice), ED (end-diastole), ES (end-systole), N (normal), HF-I (heart failure with ischemia), HF-NI (heart failure without ischemia), and HYP (hypertrophy).

• 

  Figure 6. Regression and Bland-Altman plots for LVM and EF. The left side is the LVM and EF regression curves, and the right side is the Bland-Altman plots of LVM and EF.

• 

  Figure 7. Comparison of units between the residual U-Net and our model. The solid green line represents the segmentation result, and the dotted red line represents the ground truth.

• 

  Figure 8. Segmentation results (solid green line) and ground truth (dotted red line) of three difficult slices.