A structured random inactivation UNET for retinal vascular segmentation

1. Purpose ： Reduce UNet Over fitting problem , Improve the ability of end-to-end segmentation of blood vessels

2 Main work ：

（1） suffer DropBlock Heuristic uses structured random deactivation after convolution of each layer ;





（2） Tested on three retinal image datasets SD-UNet Performance of , Namely DRIVE,STARE,CHASE_DB1;





（3） This paper SD-UNet The performance is better than UNet, And in DRIVE and CHASE-DB1 Reached on dataset SOTA.（sota It's actually State of the arts  Abbreviation , It refers to doing in a certain field Performance first-class model, Generally refers to in some benchmark Those models with very high scores on the data set ）
3, Model

 Fig 1 It shows SD-UNet Detailed structure of , You can see in the U The shape network includes three layers of lower sampling module and upper sampling module , Through the middle skip connection Connect . The lower sampling module of each layer contains two 3x3 Convolution and one-time maximum pooling , And then there was DropBlock Layer and the ReLU layer . The structure of the upper sampling part is similar , The difference is that the pooled layer is replaced by transposed convolution .

4,Dropout And Dropblock

Over fitting problem , The fundamental reason is the characteristic dimension ( Or parameter ) Too much , A perfectly trained set of functions leading to fitting , But the prediction of new data is poor .

About Dropout

In order to prevent over fitting in computer vision, deep neural networks often use different regularization methods , Random deactivation DropOut It is a simple but effective method to prevent over fitting .

The neural network contains multiple nonlinear hidden layers , This enables it to learn the complex relationship between input and output , But we can't deny that ： In practice , Even if the test set and the training set come from the same distribution , There will still be noise in the training set , Then the network will learn the distribution of data and noise at the same time , This will easily lead to over fitting .

In machine learning , We usually use model combination to improve the performance of models . However , For large neural networks , Averaging the output of multiple network models is time-consuming and space consuming . So I put forward Dropout.Dropout Is to randomly discard the neurons in the neural network , The so-called discard is to remove neurons from the neural network , Including its connection with forward and backward , Which neuron is discarded is random .

but Dropout Applicable to the whole connection layer , Does not apply to convolution , Adjacent elements in the feature graph of convolution layer share semantic information in space , The reason for this phenomenon is that convolution can extract features , With receptive field, that is to say, the network can still learn the corresponding semantic information through the adjacent location elements of the location . Therefore, the convolution network needs to be structurally and randomly inactivated , By deactivating a part of the adjacent whole area in the feature map , In this way, the network loses this part of the overall information , You can pay more attention to the learning of other parts' features .

a It's the original picture ,b The figure uses Dropout,c Use Dropblock,X Represents the discarded location . therefore DropBlock layer Contains two parameters ：block_size and γ. among block_size=1 Time and tradition dropout equally , and block_size= The whole feature map is actually Spatial Dropout;γ Parameters are responsible for controlling how many activation points in a feature map can be deactivated . The calculation formula is as follows ：

keep-prob: The probability of a neuron being preserved （ Given in advance ）

 w,h Represent the width and height of the feature graph respectively ,

block-size: Size of deactivation block .

5, experiment ：


 This section compares the systems SD-UNet With the original UNet Effect on different data sets , Will also SD-UNet With some current SOTA Compare the framework of .

Data expansion ： Random rotation 、 Color change 、 Increase Gaussian noise 、 level / Flip vertically

Performance indicators ：


PPV（ Accuracy ）,TNR（ Specificity ）,TPR（ Recall rate ）,Acc,AUC,F1score,JS（ Similarity coefficient ）

SR Represents the segmentation result ,GT It represents the actual situation on the ground .

result ：

 

UNet Indicates not used dropout Original UNet






UNet* Indicates that random deactivation is used UNet Random inactivation rate 0.25 

Table IV V VI Also shows the SD-UNet And others SOTA Comparison of methods , You can see SU-UNet Its performance is better than other methods .

  summary ：

This paper presents a method based on full convolution UNet Network structure -SD-UNet Used for pixel by pixel segmentation of retinal vessels , Mainly in the UNet add DropBlock To improve the performance of the model .
 among UNet Be responsible for capturing context information during down sampling , In the process of upsampling, different levels of features can be effectively fused ; At the same time with the help of DropBlock,SD-UNet The semantic features of some regions can be discarded during training , So as to reduce the over fitting problem .

This paper realizes SD-UNet For retinal vascular segmentation , To verify the effectiveness of this method in three open source datasets DRIVE,STARE,CHASE_DB1 Has been tested , And compared with other methods ,SD-UNet And we got it SOTA Result .