Tensorflow introductory tutorial (40) -- acunet

Today we will share Unet An improved model of ACUNet, The improved model comes from 2020 Year paper 《ACU-NET:A 3D ATTENTION CONTEXT U-NET FOR MULTIPLE SCLEROSIS LESION SEGMENTATION》, By understanding the idea of the model , stay VNet On the basis of this, we can make the same improvement .

1、ACUNet advantage

Unet Despite the success in the field of medical segmentation , But it is invalid to use context information and features to represent , It's hard to be there. MS The lesions are accurately segmented . To solve this problem , The article puts forward 3D Context attention Unet structure （ACUNet）, For segmentation MS Pathological changes , Include 3D Space attention module , In the decoding stage, it is used to enrich the spatial details and the expression of pathological features . Besides , In the decoding and encoding phase ,3D The context guidance module is used to expand the receptive field and guide local information and surrounding information .

2、ACUNet structure

2.1、3D Context boot module

MS Affected by the structure and shape of brain tissue , So the voxels around the lesion contain some information about the lesion . In order to make full use of the information around , Designed 3D Context boot module , As shown in the figure below . In this module, low dimensional local information is combined , High dimensional local information and surrounding information , The low dimensional local information is the input feature map , High dimensional local information is obtained by convolution , The surrounding information is obtained by hole convolution , The input is to combine the low dimensional local information with the high dimensional local information, and then pass through the hole ratio of 2 The result is , Finally, the output concatenates the three .

2.2、3D Spatial attention module

Spatial details are often lost in high-dimensional output , This is due to cascading convolution and nonlinearity . This makes it difficult to reduce error detection for objects with variable size and position .3D Spatial attention module can solve this problem , It produces a spatial attention coefficient on each voxel . The final output is the multiplication of input features and spatial attention coefficient elements , As shown in the figure below . To reduce the complexity of the module , Firstly, the input feature map is used with the interval of 2 The convolution operation to down sample , Reduce the resolution of the image by half . The gating vector is used to determine each voxel in the lesion area . stay ACUNet in , Input feature map x It's a low-order feature in the decoding phase , And gating vector a Represents the higher-order features in the coding phase . The whole process ：a after 3x3x3 Convolution ,x After the interval is 2 Of 3x3x3 Convolution , And then the sum of the two goes through relu Activation function combines input information and gating information , after 1x1x1 Convolution , after sigmoid Activation function , Then sample to the resolution of the original input feature map , Finally, the original input feature graph is multiplied by matrix elements to get the final output .

2.3、 Loss function

Considering the imbalance of medical images , use Focal Tversky Loss function . The best parameter in this paper is alpha=0.7,beta=0.3,gama=0.75.

2.4、3D Context attention Unet

And 3DUnet The difference , In the coding phase, we introduce 3D Context boot module , In the decoding phase 3D Context guidance module and spatial attention module .3D The context guidance module is used to expand the receptive field and guide the context information . Spatial attention module is used to enrich spatial details and expression of lesion features . Besides , In the decoding stage, the deep supervision mechanism is also introduced , It has two advantages ： Ensure that there are semantic differences in the middle layer of each scale , Make sure 3D The spatial attention module can affect the foreground content of the image .

2.5、 The evaluation index

use dice Similarity coefficient , Correctly predict the ratio , The positive rate of lesions is , The false positive rate of lesions is .

3、 Experiments and results ​

3.1、 The data used is ISBI2015 Of MS Lesion segmentation challenge data , The training set contains 5 Patients , Test set is 14 Patients . Every patient has T1,T2,PD and FLAIR Four sequence images .

3.2、 It uses GTX2080Ti The graphics card , Using random gradient descent as optimizer , The learning rate is 0.03, The attenuation parameter is 1e-6, Momentum is 0.9.MRI The image is fixed size 181x217x181, Cut to... In training 160x192x160 size . Data enhancement uses rotation and flipping .T1,T2,PD and FLAIR The combination of modal images forms the input data of four channels . Trained 80epoch, Use batch Size is 1.

3.3、 Result comparison , Compared with the existing methods ,ACUNet We can get better results .