summary

This article is an article 2020 A review of medical segmentation in , It mainly includes supervised learning and weak supervised learning , It does not include unsupervised learning . The former mainly includes the choice of backbone network 、 Design of network block and loss function , The latter mainly includes data expansion 、 Transfer learning and interactive segmentation , Focus on small samples and category imbalance . Finally, common medical segmentation data sets are introduced and some prospects are given .

details

introduction

Segmentation mainly includes semantic segmentation and instance segmentation , Semantic segmentation refers to the pixel level classification of pictures , It is to determine a category for each pixel ; Instance segmentation is based on semantic segmentation , For pixels of each category , At the same time, distinguish which instance it belongs to . And medical segmentation , The main research is semantic segmentation . And medical images have their own characteristics , Including noise and fuzzy boundaries .
Supervised learning can use sufficient labeled data for model training , But general medical image processing , There are not so many labeled images . The opposite is unsupervised learning , There is no need to annotate the data , But the model will be more difficult to train . Among them is weak supervised learning , He only needs some marked data , Most data can be unmarked , This model feels more suitable for real needs .
In addition, in terms of research , Most of the current research is data-driven , On a specific dataset , Adopt different methods to set the characteristics of data sets for design , Solve different problems .

Supervised learning

Backbone network

The backbone network in medical segmentation is basically encoder-decoder Structural , Such as FCN、U-Net、Deeplab etc. ,encoder For feature extraction ,decoder Used to reconstruct the full resolution segmentation image , Therefore, the design of backbone network also focuses on encoder More effective feature extraction and decoder Feature recovery and fusion in .
2D data The most classic backbone network in processing is UNet.
And for 3D data Of 3D net, Such as 3D-UNet etc. ,3D Data is the mainstream of medical data , however 3D Net The amount of general computing and the occupation of video memory will be large .
RNN perhaps LSTM Will also be with CNN Combined for medical segmentation tasks , Time dependence for modeling image sequences .
skip-connection operation It can fuse features of different semantic levels , At the same time, promote the spread of gradients , It is also widely used in network structures , However, feature fusion will also face the problem of too large feature gap .
Cascade of Networks , Sometimes I train 2 A or 3 A network , And cascade them , Further improve the accuracy of the network . On the task of image segmentation , It mainly includes thick and thin segmentation , Detection segmentation and hybrid segmentation . Thickness segmentation , Some practices, such as training a network for coarse-grained segmentation , Another network performs fine-grained segmentation on the basis of coarse-grained segmentation . Detect segmentation Words , First, do target detection through a detection network , Then do segmentation in the detected area . We mentioned above ,3D Net The calculation cost of is very high , So one way is to 3D The data is converted into 2D The data of , Then cascade several networks to process this part of the data . This practice is quite violent , So there is a more hot Our research is 2D-3D Mixed segmentation Methods . It's just one. 2D Network processing 2D data ,3D Network extraction 3D features , And then 2D The characteristics and 3D Feature fusion for final processing .
GAN,GAN It can be used for almost all tasks , In segmentation , One way is to generator Used to generate segmentation graph ,discriminator It is used to distinguish the generated segmentation graph from ground truth.
Prior knowledge , Using prior knowledge such as the shape and position information of organs can restrict and guide the training process of the network , So as to improve the segmentation results and improve the robustness of magic , However, there are relatively few studies on how to integrate prior knowledge into the network .