[medical segmentation] u2net

summary

u2net Is in unet Based on , It is a very effective saliency target detection model .
Significance target detection ： Segment the main body of the image .
The model proposed Main background It's two parts ：
1、 Split tasks backbone Mainly some pre training models . Because of some segmentation backbone The effect of is not as good as that of the pre training model , So we all use the pre training model . problem ： They are not designed for segmentation , Insufficient attention has been paid to the key local details and global comparison information in the segmentation task , So if you want to better adapt to segmentation, you need to add some special structures on this basis , Achieve better use of the extracted features , This brings computational complexity .
2、 The model is generally deep , Considering the overhead of video memory and Computing , We will first do some down sampling operations on the pictures in the top layers , Reduce resolution . problem ： In fact, for split tasks , The rich spatial information in high-resolution images needs to be made good use of .

u2net Successfully solved these two problems , A two-level nested u Shape structure , The network can be trained from scratch without relying on the pre training model , The effect of feature extraction is as good as the pre training model , And no additional structure is needed to deal with these features . In addition, after the network deepens , It can also maintain high resolution （ because RSU A large number of pooling operations in ）, And the overhead of video memory and computation is also very friendly .

details

Network structure

The following is u2net Network structure , In general, it's still unet Of u Shape structure , But every layer or every stage from unet The simple convolution structure in becomes RUS（Residual U-blocks）, Then each decoder The side outputs of all receive gt Supervision of , All side outputs do concat after , The convolution operation is the final network output .
Previous studies have also included stacking or cascading unet obtain $u\ast n-net$, But the author is nested or exponential unuet Went to the , Obviously, the number of nesting can be very large , namely $u^{n}net$, But considering the actual situation , Or nest one layer to get $u^{2}net$
Then each RSU The number of layers of encoder With the increase of the number of layers , namely En_1、En_2、En_3、En_4 The used are RSU-7、RSU-6、RSU-5、RSU-4, Because we attach importance to the feature extraction of high-resolution feature map , Pool operation will be performed at the same time , Reduce size . and En_5、En_6 What we use is RSU-4F 了 ,F It means that the size will not change , That is, only feature extraction .

RSU（Residual U-blocks）

RUS To replace the unet Simple convolution block in , It can better capture the overall and local information , And in the past 1x1,3x3 Convolution due to receptive field , Often good at capturing local information , The capture of global information is not so good , And the global information is often needed for segmentation .RUS Through this u The shape structure realizes the mixing of characteristic maps of different scales and different receptive fields , It can capture global information from more different scales .

And he also used the idea of residuals .resnet You need at least two levels in order to do identity mapping , Or we'll do a linear transformation , and n The linear change effect of degree is equivalent to 1 Linear transformation of degree . And in this article , Because of this u-block It contains several layers , So cross one block That's it .

Then there is the calculation quantity , The author compares some mainstream block structures , Discover though RSU The amount of calculation is linear with the increase of depth , But the coefficient is very small , Therefore, the amount of calculation is not very large , It can be stacked very deep .

Loss

There are mainly two parts , One part is the loss of side output characteristic graph , The other part is the loss of the final output feature map formed after the fusion of these side outputs .