基于3D UNet结合Transformer的肝脏及肝肿瘤自动分割

Automatic Segmentation of the Liver and Liver Tumor Based on3D UNet Combined with Transformer

目的肝脏和肝肿瘤分割是肝癌放疗计划设计的重要步骤,本文提出新型自动分割模型,以实现肝脏和肝肿瘤的精确分割。方法在3D UNet深度神经网络中加入了残差模块和Swim Transformer模块,提出一个新型的卷积和Transformer结合的Res-Swim-UNet模型。在LiTS公共数据集上对比了所提出方法与先前方法的性能,并在本地数据集上验证了Res-Swim-UNet模型的泛化能力。结果Res-Swim-UNet模型在LiTS公共数据集上肝脏分割结果的Dice相似性系数(Dice Similarity Coefficient,DSC)、体积重叠误差(Volumetric Overlap Error,VOE)分别是0.957、0.522,相对于UNet模型DSC提高了1.6%,VOE降低了1.3%;肝肿瘤分割结果的DSC、VOE分别是0.672、0.617,相对于UNet模型DSC提高了13.5%,VOE降低了5.9%。在本地数据集上肝脏分割结果的DSC、VOE分别是0.895、0.552,肝肿瘤分割结果的DSC、VOE分别是0.589、0.706。结论本文提出的Res-Swim-UNet模型可以有效提高CT图像中肝脏和肝肿瘤的分割效果,且该模型在迁移到本地数据时仍具有较高的分割精度。该模型可以用于提高医生勾画靶区的效率。

Objective The segmentation of the liver and liver tumor is an important step in the radiotherapy planning for liver cancer.In this paper,we proposed a novel automatic segmentation model to realize the accurate segmentation of liver and liver tumors.Methods The residual module and the Swim Transformer module were added to a 3D UNet deep neural network,and a new ResSwim-UNet segmentation model combining convolution and Transformer was proposed.We compared the performance of proposed and previous methods on the LiTS public dataset,and verified the generalization ability of Res-Swim-UNet model on a local dataset.Results The dice similarity coefficient(DSC)and volumetric overlap error(VOE)of the liver segmentation results of the Res-Swim-UNet model on the LiTS public dataset were 0.957 and 0.522,respectively.Compared with the UNet model,the DSC increased by 1.6%,and the VOE decreased by 1.3%.The DSC and VOE of the liver tumor segmentation results were 0.672 and 0.617,respectively,which was 13.5%higher than the UNet model in DSC,and 5.9%lower in VOE.The DSC and VOE of the liver segmentation results on the local dataset were 0.895 and 0.552,respectively,and the DSC and VOE of the liver tumor segmentation results were 0.589 and 0.706,respectively.Conclusion The Res-Swim-UNet model proposed in this paper can efectively improve the segmentation efect of liver and liver tumors in CT images,and the model still has high segmentation accuracy when transferred to local dataset.The model could be used to improve the efficiency of the target delineation for physician.