基于非对称UNet网络的磁共振图像胃肠道语义分割方法研究

Semantic Segmentation of Gastrointestinal Tract in MRI Images Based on Asymmetric Unet Network

靶区和脏器的自动化标注是磁共振图像引导放疗的关键技术之一。研究了磁共振图像中胃肠道等空腔脏器语义分割的方法,在语义分割任务中,往往输入图象比输出图像复杂很多,假设网络的复杂度和输入输出的图像复杂度正相关,提出了一个12层的非对称UNet网络,把更多的网络参数放在编码器上,解码器的参数量只有编码器的三分之一。实验结果表明在UMWGI数据集上对胃、大肠、小肠的语义分割任务中,所提方法的DSC综合得分达到了0.856,Hausdorff_95得分达到了3.743,相同网络规模的条件下,优于对称结构的UNet网络和Transfomer网络,说明所提方法可以较好地完成磁共振图像中的胃肠道语义分割,边界分割也较为理想,为实现磁共振图像上胃肠道的自动化标注提供了可行方案。

Automatic annotation of target areas and organs is one of the key technologies for MRI⁃guided radiotherapy.A method for se⁃mantic segmentation of hollow organs such as the gastrointestinal tract in magnetic resonance images is presented.In semantic segmenta⁃tion tasks,the input images are often much more complex than the output images.It is assumed that the complexity of the network is positively correlated with the complexity of the input and output images.A 12⁃layer asymmetric UNet network is proposed with more net⁃work parameters allocated to the encoder,while the decoder has only one⁃third of the parameters of the encoder.The proposed method achieves a DSC comprehensive score of 0.856 and a Hausdorff_95 score of 3.743 in the semantic segmentation task of the stomach,co⁃lon,and small intestine on the UMWGI dataset.Under the same parameter conditions,the proposed method outperforms symmetric UNet and Transformer networks,indicating that the proposed method can effectively perform semantic segmentation of the gastrointestinal tract in magnetic resonance images.The boundary segmentation is also ideal,providing a feasible solution for automated annotation of the gas⁃trointestinal tract in magnetic resonance images.