Assessment of Bi-Ventricular and Bi-Atrial Areas Using Four-Chamber Cine Cardiovascular Magnetic Resonance Imaging: Fully Automated Segmentation with a U-Net Convolutional Neural Network

Figures

• 

  Figure 1. Distribution of Dice similarity coefficient (DSC) for the four cardiac chamber segmentations using the U-Net convolutional neural network (CNN) on the test datasets. Box plots were generated by using GraphPad Prism for all cardiac phases in each chamber. (a) DSC of the left ventricle (LV), (b) DSC of the left atrium (LA), (c) DSC of the right ventricle (RV), and (d) DSC of the right atrium (RA).

• 

  Figure 2. Distribution of the Mahalanobis distance (MD) for the four cardiac chamber segmentations using U-Net convolutional neural network (CNN) on the test datasets. Box plots were generated using GraphPad Prism for all cardiac phases in each chamber. (a) MD of the left ventricle (LV), (b) MD of the left atrium (LA), (c) MD of the right ventricle (RV), and (d) MD of the right atrium (RA).

• 

  Figure 3. Four cardiac chamber segmentation—results of a representative case. Upper row: fusion images with both original image of four-chamber (4CH) cine cardiovascular magnetic resonance imaging (CMR) and label images extracted by the manual method. Lower row: fusion images with both the original image of 4CH cine CMR and label images extracted by the fully automated method with U-Net convolutional neural network (CNN). Left column: fusion images at the end-diastolic phase. Right column: fusion images at the end-systolic phase. RV = right ventricle; LV = left ventricle; RA = right atrium; LA = left atrium; DSC = Dice similarity coefficient.

• 

  Figure 4. Correlation (left) and Bland-Altman (right) plots of the end-diastolic area (EDA) in each chamber generated from the manual method and the U-Net CNN. (a) Left ventricular (LV) EDA, (b) left atrial (LA) EDA, (c) right ventricular (RV) EDA, and (d) right atrial (RA) EDA.

• 

  Figure 5. Correlation (left) and Bland-Altman (right) plots of the end-systolic area (ESA) in each chamber generated from the manual method and U-Net CNN. (a) Left ventricular (LV) ESA, (b) left atrial (LA) ESA, (c) right ventricular (RV) ESA, and (d) right atrial (RA) ESA.

• 

  Figure 6. Correlation (left) and Bland-Altman (right) plots of fractional area change (FAC) in each chamber generated from the manual method and U-Net CNN. (a) Left ventricular (LV) FAC, (b) left atrial (LA) FAC, (c) right ventricular (RV) FAC, and (d) right atrial (RA) FAC.

• 

  Figure 7. Receiver operating characteristic (ROC) curves of the chamber area extracted by the U-Net convolutional neural network (CNN) for identifying a patient with an enlargement of each heart chamber. (a) Left ventricle (LV), (b) left atrium (LA), (c) right ventricle (RV), and (d) right atrium (RA). AUC = area under the curve.