Tensorflow introductory tutorial (37) -- DC Vnet

Today we will share Unet An improved model of DC-UNet, The improved model comes from 2020 Year paper 《DC-UNet Rethinking the U-Net Architecture with Dual Channel Efficient CNN for Medical Images Segmentation》, By understanding the idea of the model , stay VNet On the basis of this, we can make the same improvement .

One 、 original Unet The Internet ​

1、  original Unet There are some flaws . Therefore, the author of the article puts forward some improvements Unet Of DC-Unet（ Two channels Unet）, It is mainly designed from two aspects , One is to design an efficient CNN Structure takes the place of Unet Encoder and decoder in , The second is to replace the hopping connection between encoder and decoder with residual module . Relative to the original Unet result , On the three data, the results were improved 2.9%,1.49% and 11.42% The accuracy of the .

2、 Unet Classic structure

By convolution , Pooling layer , Upper sampling layer , Jump connection layer composition .

Two 、 MultiResUNet The Internet ​

1.1、 Use convolution kernels of different sizes to replace the original Unet Convolution to extract features , Then the results of different convolutions are spliced and connected , Realize the feature fusion of different image scales . As shown in the figure below , In order to reduce the number of channels , Thus reducing parameter calculation , adopt 1x1 Convolution to reduce dimensions , As shown in the figure below .

In the paper, the author made a comment on the above Inception Module simplified , Through a series of the same convolution operation combination, the calculation effect of convolution kernels of different sizes can be achieved , Then different convolution results are stitched to complete the stitching and fusion of different scale features , Here's the picture b Shown . Besides , The author also adds residuals connected to Inception Go to the module , Here's the picture c Shown .

2.2、Unet Replace the jump connection module with residual path connection , As shown in the figure below , The input of this module is the output of the decoding network , After four 3x3 Residual connection convolution operation , Finally, it is spliced with the output of the coding network .

2.3、 With multi-scale residual module and residual path MultiResUnet The structure is as follows . The size and number of convolution kernels of each multiscale residual network is 3 individual 3x3 and 1 individual 1x1, The number of characteristic channels is different for each module , See the table for details 1. The convolution kernel size of the residual path is multiple 3x3 and 1x1 Superposition of modules , You can see from the figure below that 3x3 and 1x1 The number of residual modules of are 4,3,2,1. See the table for details 1.

3、 ... and 、 DC-UNet The Internet ​

in the light of MultiResUnet The model proposes a more effective network structure to extract more spatial features . In order to overcome the problem of invalid spatial features , Using multiple channels 3 individual 3x3 Convolution layer instead of MultiRes Residual connection in the module . The author of this paper adopts 2 A passage 3 individual 3x3 Convolution layer , Finally, the results of these two channels are linearly added and output . As shown in the figure below .

With dual channel module and residual path DC-Unet The structure is as follows . The convolution kernel size of each two channel module is 3x3 size , The number of characteristic channels is different for each module , See the table for details 2. The convolution kernel size of the residual path is multiple 3x3 and 1x1 Superposition of modules , You can see from the figure below that 3x3 and 1x1 The number of residual modules of are 4,3,2,1. See the table for details 2.

​

Four 、  Training details ​

Preprocessing unifies the image to 256x128 size , And change the image from 16 The bit data format is converted to 8 Bit data format . Using the binary cross entropy loss function , The optimizer uses Adam And its default parameters .

5、 ... and 、  The evaluation index ​

Because the output image is 0 To 1 The numerical , You can use the binary comparison method ：jaccard Similarity is IOU. You can use the gray comparison method ： Mean error of absolute value （MAE）,Tanimoto similarity （jaccard Extended derivation of similarity ）, Structural similarity （SSIM）.

6、 ... and 、  Results contrast ​

Compared on three data sets UNet,MultiResUNet and DC-UNet, No matter from the objective evaluation index or the supervisor's vision ,DC-Unet The results are the best .