Cross-Domain Echocardiography Segmentation with Multi-Space Joint Adaptation

Figures

• 

  Figure 1. Echocardiography samples from different medical centers and vendors (a) CAMUS dataset: GE Vivid E95 ultrasound scanners with a GE M5 S probe and (b) self-acquired dataset: Vinno G86 ultrasound scanners with a S1–6 PX probe.

• 

  Figure 2. Overview of the proposed MACS. The MACS consists of an attention-based dual-branch segmentation module (ASDB), a discriminator in feature space, and a discriminator in the output space. Source domain images (XS) and target domain images (XT ) are segmented by the ASDB which is optimized using only source domain samples. The output of the segmentation network contains predicted masks (PXm ) and high-level intermediate features (FX ). The entropy map (EX) is calculated from the segmentation prediction PXm. and EX from two different domains are forced to be aligned by discriminators, making the ASDB output accurate segmentation on unseen target domain dataset.

• 

  Figure 3. The architecture of the Attention Segmentation with the Dual-Branch (ASDB) module. The encoder extracts multiply single-layer feature maps. The multi-layer feature is the fusion of these feature maps. Single-layer features and multi-layer feature pass by the deep attention model (DAM) for the refined feature. The dual-branch structure generates the prediction of the mask and boundary.

• 

  Figure 4. Representative visual segmentation results of the ablation study. Different model configurations are compared, including the ASDB, the ASDB + Feature discriminator (Df), the ASDB + Entropy discriminator (De), and the full framework of the MACS. The green and blue contours indicate boundaries of the LVendo and LVepi, respectively. The yellow and orange colors denote the ground truth.

• 

  Figure 5. The mean Dice Index (DI) for segmentation results of the MACS trained with different numbers of target domain images. (a) The line graph of the mean DI. (b) The boxplot of the mean DI.

• 

  Figure 6. Fake images generated by image-to-image translation methods (SAGAN, CUT, F-LSaSim) and the visualization prediction of fake images that segmented by the DAF.

• 

  Figure 7. The Bland-Altman analysis on adversarial learning methods about the organ area (×103 pixel2). The x-axis represents the average area of the prediction and the ground truth (GT). The y-axis represents the bias of the area of the prediction and the GT. The forest green or royal blue line indicates the bias of the organ area and brown lines indicate 95% confidence intervals of the area bias.

• 

  Figure 8. Representative visual segmentation results of adversarial learning methods. Green and blue contours indicate boundaries of the LVendo and LVepi, respectively. Yellow and orange colors denote the ground truth.