BBTUNet:基于上下文Transformer的肝脏肿瘤分割算法研究  

BBTUNet:Research on liver tumor segmentation algorithm based on context Transformer

作　　者：宋长明[1] 宋蒙 肖露 梁朝阳 彩朔 SONG Changming;SONG Meng;XIAO Lu;LIANG Chaoyang;CAI Shuo(School of Science,Zhongyuan University of Technology,Zhengzhou 451191,China)

机构地区：[1]中原工学院理学院,河南郑州451191

出　　处：《电子设计工程》2024年第5期190-195,共6页Electronic Design Engineering

摘　　要：肝癌是世界范围内最常见的恶性肿瘤之一,严重威胁着人类的生命健康,从计算机断层扫描(Computed Tomography,CT)中精确分割出肝脏肿瘤对后期的临床诊断具有重要的意义。现有的方法虽然实现了肝脏肿瘤的自动化分割,但肝脏肿瘤边界模糊、目标较小、容易漏检等问题尚未很好地解决,肝脏肿瘤的精确分割仍旧是一项极具挑战的任务。针对这些问题,该文提出一种新的分割网络BBTUNet。构建基于Transformer的上下文Bridge,重新设计UNet的跳跃连接结构,有效捕捉多尺度特征之间的上下文关系。在Transformer的前馈神经网络中引入可分离的空洞卷积,提出改进的前馈神经网络BFFN,有效融合全局和局部信息,增强边界特征,细化分割边缘。在3DIRCADB数据集上对模型进行训练和测试,实验结果表明,提出的BBTUNet网络的Dice系数为82.1%,ACC为96.4%,相较于UNet网络,分别提升了10.9%、4.6%,且对于小尺寸、低对比度、边界模糊的肿瘤分割具有显著优势。Liver cancer is one of the most common malignant tumors in the world,which seriously threatens human life and health.Accurate segmentation of liver tumors from Computed Tomography(CT)is of great significance for later clinical diagnosis.Although the existing methods have realized the automatic segmentation of liver tumors,the problems such as fuzzy boundary of liver tumors,small target,easy to miss detection and so on have not been solved well.The accurate segmentation of liver tumors is still a very challenging task.To solve these problems,this paper proposes a new partition network BBTUNet.Build a context Bridge based on Transformer,redesign the jumping connection structure of UNet,and effectively capture the context relationship between multi-scale features.Separable hole convolution is introduced into Transformer’s feedforward neural network,and an improved feedforward neural network BFFN is proposed,which effectively fuses global and local information,enhances boundary features,and refines segmentation edges.The model is trained and tested on the 3DIRCADB dataset.The experimental results show that the Dice of the proposed BBTUNet network is 82.1%,and the ACC is 96.4%,which is 10.9%and 4.6%higher than that of the UNet network respectively,and it has significant advantages for tumor segmentation with small size,low contrast and fuzzy boundary.

关 键 词：肝肿瘤分割 UNet TRANSFORMER 上下文Bridge 

分 类 号：TN919.81[电子电信—通信与信息系统]