[paper reading] cavemix: a simple data augmentation method for brain vision segmentation

[ Address of thesis ][ Code ][MICCAI 21]

Abstract

Segmentation of brain lesions (Brain Lesion Segmentation) It provides a valuable tool for clinical diagnosis , Convolutional neural networks (CNN) We have achieved unprecedented success in this task . Data enhancement is a widely used strategy , Can improve CNN Training effect , The design of enhancement methods for brain lesion segmentation is still an open problem . In this work , We propose a simple data enhancement method , go by the name of CarveMix, Used based on CNN Segmentation of brain lesions . With others based on " blend " The same way , Such as Mixup and CutMix,CarveMix Randomly combine two existing marker images to generate new marker samples . However , Different from these image combination based enhancement strategies ,CarveMix It is disease perception , Pay attention to pathological changes when combining , And create appropriate annotations for the generated image . say concretely , According to the location and geometry of the lesion , Carve a region of interest from a marked image (ROI),ROI The size of is sampled from a probability distribution . then , Carved ROI It replaces the corresponding voxel in the second marked image , The annotation of the second image is also replaced accordingly . In this way , We generate new tagged images for network training , And the lesion information is preserved . In order to evaluate the proposed method , We conducted experiments on two brain lesion datasets . It turns out that , Compared with other simple data enhancement methods , Our method improves the accuracy of segmentation .

Method

This paper is a data enhancement method specially proposed for brain lesion segmentation ——CarveMix. This method is also a method based on label fusion , such as MixUp(ICLR 18) Is to fuse the two labels linearly , and CarveMix(ICCV 19) It is a kind of nonlinear fusion . It should be noted that , These classical methods are used for image classification tasks , Therefore, there is a lack of label fusion methods for segmentation tasks .CarveMix The integration process of is as follows ：

First look at the formula directly ： Fused image $\mathbf{X}$ And the label obtained by fusion $\mathbf{Y}$ The final calculation process of is as follows ：$$\mathbf{X}={\mathbf{X}}_i\odot {\mathbf{M}}_i+{\mathbf{X}}_j\odot \left(1-{\mathbf{M}}_i\right)$$$$\mathbf{Y}={\mathbf{Y}}_i\odot {\mathbf{M}}_i+{\mathbf{Y}}_j\odot \left(1-{\mathbf{M}}_i\right)$$ For the sake of intuition , We take label fusion as an example to show the specific fusion process , The fusion of image itself is consistent with the fusion of label . Seen from the figure , The fusion of labels is basically equivalent to directly merging two original images mask ${\mathbf{Y}}_i$ and ${\mathbf{Y}}_j$ Add it up directly ：

In fact, it is ${\mathbf{Y}}_i$ Multiplied by a factor $a$ After and ${\mathbf{Y}}_j$ Multiplied by a factor $b$ And then add up , Yes $a+b=1$. In the figure "⊙" The symbol represents pixel by pixel , So you can even put ${\mathbf{M}}_i$ As a spatial attention map .

The problem now is actually how to calculate ${\mathbf{M}}_i$ 了 . As you can see from the diagram ,${\mathbf{M}}_i$ And ${\mathbf{Y}}_i$ In fact, it is very similar , It's a bit like in ${\mathbf{M}}_i$ An expansion operation is carried out on the basis of . say concretely ,${\mathbf{M}}_i$ pass the civil examinations $j$ A pixel value ${\mathbf{M}}_i^v$ The calculation method is as follows ：$${\mathbf{M}}_i^v=\{\begin{array}{l}1,D^v\left({\mathbf{Y}}_i\right)\le \lambda \\ 0,\text{ otherwise }\end{array}$$$$D^v\left({\mathbf{Y}}_i\right)=\{\begin{array}{rl}-d\left(v,\partial {\mathbf{Y}}_i\right),& \text{ if }{\mathbf{Y}}_i^v=1\\ d\left(v,\partial {\mathbf{Y}}_i\right),& \text{ if }{\mathbf{Y}}_i^v=0\end{array}$$ This $d(v,\partial {\mathbf{Y}}_i)$ Refers to the current pixel $v$ And " Lesion boundary "$\partial {\mathbf{Y}}_i$ Distance of . You can see , If $v$ Itself is in the lesion area , that $D^v\left({\mathbf{Y}}_i\right)$ Just give a negative number , So as to ensure that it has a greater probability of being selected ( That is less than $\lambda$); And if it is outside the lesion area , We think that the closer we are, the more we should be selected . It is worth noting that , This $\lambda$ It's positive and negative , So as to ensure the pathological area " inflation " perhaps " shrinkage ".$\lambda$ The specific calculation process of is more complex , Interested readers can read the original .