An improved contrastive learning network for semi-supervised multi-structure segmentation in echocardiography

Figure 1. Overview of the proposed framework. The proposed framework can be divided into two branches: supervised and unsupervised. The supervised branch was trained with a few epochs firstly, followed by training the unsupervised branch with a lot of epochs. The supervised branch consists of a convolutional neural network (CNN) to extract the feature map of labeled images and a classifier to make predictions. In the unsupervised branch, the same CNN and classifier from the supervised branch were used, the classifiers in two branches share parameters with each other. In CNN, each blue box represents a multi-channel feature map and each white box represents a copied feature map. The downward red arrows indicate the downsampling stages and the upward red arrows indicate the upsampling stages. The blue arrow to the right represents a convolution of 1×1 and the black arrow to the right indicates a skip connection. The dilated convolutions were implemented in the downsampling stages. Furthermore, a projector was introduced to modify the feature dimensions. In the supervised branch, the loss function employed a standard cross-entropy loss (Lce), which measures the discrepancy between the…

Figure 2. Performance comparison of our ablation studies on 2CH images from CAMUS dataset. (A) The IoU values were examined to assess the performance of our ablation studies. (B) The validation loss values were analyzed to evaluate the performance of our ablation studies.

Figure 3. The variability of IoU about our ablation studies. (A) The results obtained from 2CH images. (B) The results obtained from 4CH images.

Figure 4. The Boxplots show the IoU values for both 2CH and 4CH images, using the four methods mentioned in Table 2. These boxplots specifically showcase the outcomes achieved when employing 1/2 labeled images. (A) The results obtained from 2CH images. (B) The results obtained from 4CH images.

Figure 5. The IoU for segmentation results of the proposed method trained with different numbers of labeled images. (A) The results obtained from 2CH images. (B) The results obtained from 4CH images.

Figure 6. Visual comparison of CAC method and the proposed method with 1/4labeled, 1/2labeled and full labeled data on 2CH and 4CH images. The red, green and blue lines represent the ground truth. The value of IoU is marked on the right upper corner of each image. (A,B) The visual comparison of the performance about discriminating the borders of myocardium on 2CH images and 4CH images. (C,D) The visual comparison of the probability of segmentation errors in some regions on 2CH images and 4CH images. (E,F) The visual comparison of the performance about the identification of the ventricle and myocardium location.

Figure 7. Visual comparison of CAC method and the proposed method on 6 typical challenging images. The red, green and blue lines represent the ground truth. The value of IoU is marked on the right upper corner of each image. (A–C) The visual comparison of the performance on some typical low contrast images. (D) The visual comparison of the performance on images where the complete cardiac structures are not present. (E,F) The visual comparison of the performance on images that have no a clear border between the ventricle and the atrium.