[medical segmentation] unet3+

summary

unet3+ Is in unet as well as unet++ Based on ,unet The core is skip-connection.
and unet++ It is also the change made on this core , Overlapping dense convolution is used instead of rough feature fusion .
and unet3+ be aware ,unet and unet++ Not extracting enough information directly from multi-scale information , Based on this , Designed a new skip-connection Structure , Better integrate low-level fine-grained information and high-level semantic features , And the parameters of this structure will be larger than unet and unet++ Few . meanwhile , about decoder In depth monitoring of the output of , A new loss function is proposed for training . On the other hand , Use classification as a guide , Reduce over segmentation in the background image .
notes ：unet and unet++ Not extracting enough information directly from multi-scale information refer to ,unet According to the corresponding layer encoder And the next layer decoder The result of up sampling constructs the current layer decoder Of , There is no use of multi-scale information ; and unet++ Although the use of multi-scale information is carried out through nesting and dense skip connection , But more like for encoder Continuous processing of features , Not the use of original features ;unet3+ It is the use of multi-scale information of original features .
The following is the comparison diagram of the three structures ：

details

Full-scale Skip Connections

unet3+ Still unet Based on the changes , So he is still encoder-decoder Structure .encoder Part of it is actually unet,unet++ also unet3+ It's all the same , The key is decoder How did you get the part .unet3+ The way to do this is ,encoder After pooling and convolution operations, the characteristic map with the number of middle layers less than or equal to the current layer is obtained 64 Characteristic graph of channels （ Of course , The same layer does not require pooling ）, then decoder The feature map with the number of middle layers greater than the current layer is up sampled （ linear interpolation ） Same as convolution 64 Characteristic graph of channels , Then these characteristic figures concat get up , In the form of 5 Layer structure as an example , There will be a total 64x5=320, this 320 The characteristic graph of the number of channels constitutes decoder The first floor of .
The formal expression of the above process is ：

also , One thing I have to mention is , Even though unet3+ The structure of is relative to unet It will be complicated , however encoder and decoder Structurally , On the contrary, the number of parameters will be less . That is to say encoder and decoder Structurally , The rank of parameter quantity is ：unet3+<unet<unet++

Full-scale DeepSupervision

unet3+ and unet++ All adopt in-depth supervision , I think it's almost the same …unet++ With this in-depth supervision, we can achieve the effect of additional model pruning , and unet3+ It seems that there is only basic effect ？

Loss function ： yes MS-SSIM loss+focal loss+ IoU loss

Classification guided segmentation

stay encoder The last layer or decoder The last floor of , Then a classifier , Determine whether the picture contains the target , This output will participate in the segmentation , Do joint training , Reduce the segmentation of background pictures .