MUF-Net: A Novel Self-Attention Based Dual-Task Learning Approach for Automatic Left Ventricle Segmentation in Echocardiography

Figures

• 

  Figure 1. Network structure diagram. The top half represents the segmentation branch, and the bottom half represents the optical flow branch.

• 

  Figure 2. The architecture of the proposed MSEA-U-Net. The blue block indicates the deep semantic multi-scale fusion module (DSMSFM), and the purple block indicates the module-edge position-attention module (EPAM).

• 

  Figure 3. Deep Semantic Multi-scale Fusion Module (DSMSFM). The bottom is the overall structure of DSMSFM, and the top is the details of the self-attention module.

• 

  Figure 4. The architecture of Edge Location Attention Module (ELAM). The blue part and the red part are the Location attention branch, and the pink part is the Information attention branch.

• 

  Figure 5. The architecture of the modified FlowNet (mFlowNet). The blue rectangles represent the original FlowNet blocks, while the green rectangles represent the modified parts by ours.

• 

  Figure 6. The segmentation results of the ES frame and ED frame. From (a–c), each column is a sample of the ES and ED frames in the video. The green line is the result of this work. The blue line is the result of the contrast algorithm (JSAO-Net), and the red line is the label. The white box presents the main difference between the segmentation results of the two algorithms.

• 

  Figure 7. The comparison results of unlabeled transition frames are shown in the figure. (a,f) are ES and ED frames, respectively, and (b–e) are four transition frames between them. The left side of each image shows the original image and the right side shows the corresponding comparison visualization. The green line is the result of this work. The blue line is the result of the JSAO-Net, and the red line is the label. The white box presents the main difference between the segmentation results of the two algorithms.