Interpretation of the paper: attention mechanism in medical images

Click on the above “ Xiaobai studies vision ”, Optional plus " Star standard " or “ Roof placement ”

 Heavy dry goods , First time delivery 

source ｜Daniel [email protected] You know ,https://zhuanlan.zhihu.com/p/138555896

Multi-scale self-guided attention for medical image segmentation

The method of this paper is 2019 Published in CHAOS MRI Dataset The best method of medical image segmentation on , The final Dice by 86.75. 

Introduction： Views on previous medical image segmentation methods

 

One 、 Views on past models

1、 In most classical models, the use of multiple sizes , Such as unet structure ,FCN structure . Because at the beginning Again Of low-level Continuous feature extraction of information , So it will cause information redundancy Use .

2、 When the past model is applied to the pixel level segmentation challenge （ Such as , Segmentation of the medical field ）, May show a lack of discrimination .

Two 、 For the current methods used to improve the expression ability of learning features , Such as multi-scale context fusion , Using hole convolution ,pooling And so on

1、 Although the previous practice can obtain the information of the target in different sizes , But for all image The contextual connections are homogenous And non adaptive , Ignore In different categories ,local-feature And context dependency .

2、 These multi-scale context dependencies are basically artificially set , Lack of flexibility of the model itself . This makes it impossible to make full use of some long-distance connections in the image , These are very important for medical image segmentation .

3、 ... and 、 Views on the mechanism of attention

1、 The application of attention mechanism can well characterize the synaptic segmentation region , And suppress other noise parts .

2、 The author also uses a relatively perfect attention mechanism in this paper ：a. Use attention mechanism for different sizes of semantic information ;b. Each attention model consists of position Attention model and channel The attention model consists of two parts ; In this way, the model can learn more extensive and rich context dependent information , And different promotion channel The degree of dependency between contexts .

Method： A detailed description of the thesis model

One 、 General description of the model

Because in the conventional CNN in ,global feature It's usually based on local receptive field To obtain the . So for long-range The context dependence of is not fully expressed . Therefore, the author introduces the attention mechanism to solve this problem . First, obtain the overall situation through the multi-dimensional strategy feature, Then the learned global features are introduced into the attention module . Attention module consists of spatial attention module and channel The attention module consists of two parts . On the one hand, the attention module can help the fusion of local features and global features , On the other hand, it can also filter irrelevant noise information . The overall structure of the model is shown in the figure below ：

Two 、 Multi dimensional attention mechanism

The whole model is based on ResNet-101 Improvements made ,Res-2, Res-3, Res-4, Res-5 The generated characteristic graph F0,F1,F2, F3; They are sampled to the same size by linear interpolation , namely F's. Talk about the generated F'0, F'1, F'2, F'3 Conduct concate operation , Then perform convolution , Generate Multi Size fusion feature map FMS:

next , Generative FMS And F'0, F'1, F'2, F'3 separately concate operation , Then convolution , Last fed into Attention model . Get the attention characteristic map under different sizes A0, A1,A2,A3:

3、 ... and 、 Space and channel Attention map

In this article spatial and channel attention self-attention modules Mainly from another paper ：Dual Attention Network for Scene Segmentation

position attention module part： Used to capture long-distance dependencies , Solve the problem of local receptive field .

Of the first three branches , The first two branches F0 and F1 Calculate the correlation matrix between positions

Then by the correlation matrix between positions

Guide the third branch

Calculate the spatial attention map , Weighted sum with input ：

channel attention module(CAM): Capture the dependencies between channels , Enhance the representation of specific semantic features .

And spatial attention module similar , Of the three branches , The first two branches are calculated first channel Correlation matrix of different positions in , Multiply by the third branch , Get about channel My attention map , Finally, the weighted sum with the input ：

Finally, spatial attention map and channel attention map are added to get space - Channel attention signature .

Four 、 attention loss

By setting up different loss, To guide the model in training , Self adjustment requires attention , The attention mechanism model of a certain size is shown below ：

F On the one hand, input it into the attention model to generate the attention feature map , On the other hand, enter a unet In structure , Two encode Part production first attention loss：

Two unet The structure of the structure output generates a second attention loss:

Division loss by ：

Final total loss by ：

The good news ！

Xiaobai learns visual knowledge about the planet

Open to the outside world

 download 1：OpenCV-Contrib Chinese version of extension module 

 stay 「 Xiaobai studies vision 」 Official account back office reply ： Extension module Chinese course , You can download the first copy of the whole network OpenCV Extension module tutorial Chinese version , Cover expansion module installation 、SFM Algorithm 、 Stereo vision 、 Target tracking 、 Biological vision 、 Super resolution processing and other more than 20 chapters .


 download 2：Python Visual combat project 52 speak 
 stay 「 Xiaobai studies vision 」 Official account back office reply ：Python Visual combat project , You can download, including image segmentation 、 Mask detection 、 Lane line detection 、 Vehicle count 、 Add Eyeliner 、 License plate recognition 、 Character recognition 、 Emotional tests 、 Text content extraction 、 Face recognition, etc 31 A visual combat project , Help fast school computer vision .


 download 3：OpenCV Actual project 20 speak 
 stay 「 Xiaobai studies vision 」 Official account back office reply ：OpenCV Actual project 20 speak , You can download the 20 Based on OpenCV Realization 20 A real project , Realization OpenCV Learn advanced .


 Communication group 

 Welcome to join the official account reader group to communicate with your colleagues , There are SLAM、 3 d visual 、 sensor 、 Autopilot 、 Computational photography 、 testing 、 Division 、 distinguish 、 Medical imaging 、GAN、 Wechat groups such as algorithm competition （ It will be subdivided gradually in the future ）, Please scan the following micro signal clustering , remarks ：” nickname + School / company + Research direction “, for example ：” Zhang San  +  Shanghai Jiaotong University  +  Vision SLAM“. Please note... According to the format , Otherwise, it will not pass . After successful addition, they will be invited to relevant wechat groups according to the research direction . Please do not send ads in the group , Or you'll be invited out , Thanks for your understanding ~