A Deep Learning Segmentation Approach in Free-Breathing Real-Time Cardiac Magnetic Resonance Imaging

Figure 1. Comparison of breath-hold CMR data (a, c) and free-breath CMR data (b, d), where (a, b) are ED frames and (c, d) are ES frames.

Figure 2. The block diagram of the proposed method for segmentation.

Figure 3. The proposed network architecture for heart segmentation. The label value of LV, LV myocardium, RV, and background is 4, 3, 2, and 1, respectively.

Figure 4. Original images in ACDC data (a, c) and corresponding labeled images (b, d).

Figure 5. The process of data augmentation in training.

Figure 6. Comparison of the learning curves of proposed network and U-Net with cross-entropy (CE) loss, improved generalized Dice (IGD) loss, and data augmentation (DA). Dice curve (a, c) and loss curve (b, d) of training and validation sets, respectively. Note that the learning curves have been smoothed by moving average method.

Figure 7. Representative ED and ES frames segmentation results by proposed method from slice 1 to slice 5 in short axis. Red, green, and blue indicate LV, LV myocardium, and RV, respectively. Note that the results have been cropped for better observation of the heart region (GT: ground truth).

Figure 8. Representative ED and ES frames segmentation results by the proposed method from slice 6 to slice 10 in short axis. Red, green, and blue indicate LV, LV myocardium, and RV, respectively (GT: ground truth).

Figure 9. Failure cases from the heart segmentation model.

Figure 10. Segmentation results of six cycles at ED frames in different slices. The first row is near the apex slice, the second row is mid-slice, and the third row is near basal slice.