-
Article transferred from wechat official account: “Machine Learning Alchemy”
-
Author: Brother Lian Dan (authorized)
-
Contact: wechat CYX645016617 (welcome to exchange and make progress together)
-
Title: “Dual Network Generative Adversarial Networks for Pediatric Echocardiography Segmentation”
0 Preparations
0.1 the new words
- Pediatric children
- Pediatric echocardiography
- Congenital heart disease (CHD
0.2
1 review
In order to obtain high-quality segmentation results, the segmentation of pediatric echocardiography is mainly completed by ultrasonic workers manually, which is time-consuming and laborious, and highly dependent on the professional level of ultrasonic workers. In order to solve these problems, a new convolutional neural network (CNN) structure is proposed, which is called dual network General Adversarial Network (DNGAN). DNGAN consists of a generator and two discriminators. The generator uses a parallel dual network to extract more useful features to improve performance. We use a double discriminator to force the generator to learn more spatial features and segment the edge of the left heart more accurately.
2. Problem Presentation
3 Model Structure
The structure of this DNGAN is shown above: it contains a generator and two discriminators.
3.1 the generator
It is composed of a U-Net and FCN in parallel. Two features are extracted from the input images respectively, and then the features are multiplied by pixels.
The output characteristic graph of FCN is F1F_1F1, the output characteristic graph of U-NET is F2F_2F2, then the image segmentation result output by the generator is FG= F1 ×f2F_G = F_1\ times f_2FG= F1 × F2.
3.2 discriminator
Is a six-layer full convolutional network, and then use 7, 5, and 3 as the size of the convolution kernel respectively. The convolution layer is followed by the BN layer and LeakReLU activation layer.
The multi-scale L1 loss is used, similar to the 2014 Richer Conv Net image segmentation network.
4 loss
4.1 the generator losses
Let’s review the loss function of GAN in general:
- X is the real image
- Z is random import for generator
- G(z) is the generated mask
- D of x is the probability that x is true
Let’s look at the loss function of DNGAN:
- N is the number of samples;
- Xnx_nxn is a four-chamber echocardiogram of a child, and YNy_NYn is the corresponding ground truth.
- Lml_mlm is mean absolute error, also known as L1loss;
Is the discrimintor feature. L indicates the number of discrimintor layers
Indicates the discrimintor feature extracted at layer I
- Lcosl_ {cos}lcos is known as the cross entropy;
- LD2l_ {D_2}lD2 is the second discriminitor loss function. Let’s see, this loss function discriminates between a generated mask and a real one.
5 Data Description
- The dataset contained 87 echocardiograms of children;
- They collected healthy children aged 0 to 10, and each video contained at least 24 frames and a completed cardiac cycle;
- 67 videos were randomly selected and 1765 images were selected as training sets. 451 videos were selected from the remaining 20 videos as the test set;
- The resolution of the original image is 1016×708 or 636×432. All the images will become 704×704 and 448×448 after passing through the crop center.
6 summarizes
1. There are some doubts about the results and process of this article. There is a sign error of a formula in the article.
(2) The article does not give the reason why loss of the second discrimintor uses the average value of MSE and BCE. It is uncertain whether the advantages of this method have been discussed in previous articles. According to the results, the improvement caused by using GAN’s framework for training is not high, but the focus of improving segmentation accuracy should be the feature fusion and width increase of segmentation network. 3. The application of GAN in children’s ultrasound heart disease has a good starting point, hoping to help more children get rid of the trouble of congenital heart disease.