一种基于TransUNet的双分支并行医学图像分割模型

A Two-Branch Parallel Medical Image Segmentation Model Based on TransUNet

摘要: 医学图像分割任务中，为充分利用TransUNet模型能有效捕获全局和局部特征的优势，在其基础上，提出DouTransNet模型，编码器部分针对单分支Transformer模块学习角度单一、容易丢失细节特征的问题，将Transformer设计成双分支并行结构，来提取不同尺度的特征，融合两个分支的特征，实现特征互补；针对在融合两个分支的特征时可能存在冗余信息问题，添加多核并行池化模块，在保留多尺度特征的同时去除冗余信息；在解码器设计多尺度融合模块(USF)，融合来自编码器的三个尺度信息，有效弥补编码器与解码器之间的信息差距。在Synapse和ACDC数据集上进行了多次对比实验，Synapse数据集上平均DSC系数可达79.20%，较TransUNet模型提高3.34%，HD距离为25.24%，降低了11.67%；在ACDC数据集上平均DSC系数可达90.30%，提高1.67%。

Abstract: In the medical image segmentation task, in order to make full use of the advantages of TransUNet model which can effectively capture global and local features, the DouTransNet model is proposed based on it. The encoder part aims at the problems of single learning Angle and easy loss of detail features in single-branch Transformer module. Transformer is designed as a two-branch parallel structure to extract features of different scales and fuse features of two branches to achieve feature complementarity. To solve the problem of redundant information when fusing the features of two branches, a multi-core parallel pooling module is added to remove redundant information while retaining multi-scale features. In the decoder design multi-scale fusion module (USF), the information of three scales from the encoder is fused to effectively bridge the information gap between the encoder and the decoder. Several comparison experiments were conducted on Synapse and ACDC data sets. The average DSC coefficient on Synapse data set can reach 79.20%, which is 3.34% higher than TransUNet model, and the HD distance is 25.24%, which is 11.67% lower. The average DSC coefficient on ACDC dataset can reach 90.30%, an increase of 1.67%.