[reading of the paper] a multi branch hybrid transformer network for channel terminal cell segmentation

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

Abstract

Corneal endothelial cell segmentation is used to quantify cell density 、 Coefficient of variation and hexagon and other clinical indicators play an important role . However , The uneven reflection of corneal endothelium and the trembling and movement of the subject cause the edge of cells in the image to be blurred , It's hard to separate , More details and background information are needed to release this problem . Because the acceptance domain of local convolution and continuous down sampling is limited , The existing deep learning segmentation methods can not make full use of the global background , Missed a lot of details . This paper proposes a new method based on Transformer Mixed with multiple branches of trunk edge branches Transformer The Internet (MBT-Net). First , We use convolution blocks to focus on local texture feature extraction , And pass Transformer Establish space with residual connection 、 Long range dependence of channels and layers . Besides , We use trunk edge branches to promote local consistency and provide edge location information . In the data set collected TM-EM3000 And public Alisarine On dataset , With other advanced methods (SOTA) comparison , The proposed method achieves improvement .

Method

The task of this paper is corneal endothelial cell segmentation . The proposed network is less about using Transformer, It's better to put several Transformer Layer is used as an advanced global context feature enhancement tool . As you can see from the diagram , In this paper, the Encoder-Decoder The architecture is very simple , First, a convolution layer is used to extract the features of the shallow layer , And send it into Transformer Extract deep features from layers .

Another two points of this article should refer to the use of "Edge Supervision", To improve the accuracy of segmentation results at the edge of cells . And edge segmentation ground truth It's also easy to get , Use it directly canny Operator pair primitive ground truth Just extract it .

Experiment

And LinkNet34,DinkNet34,UNet,UNet++,TransUNet Made a comparison ：

In fact, strictly speaking, the method of comparison is not the method of doing this cell segmentation task , It's a relatively small direction .