【作者】 张军；

【导师】 梁栋； 徐凯； 夏柱海；

【作者基本信息】 安徽大学 ， 新一代电子信息技术（含量子技术等）（专业学位）， 2023， 硕士

【摘要】 在医学图像处理领域中,自动勾画食管癌靶区是一个备受关注的问题。确定靶区范围对于放射治疗的有效性和毒副作用的发生率有直接影响。传统的手动勾画方法需要大量时间和人力,且易受到医生主观判断和技能水平的影响,因此自动化靶区勾画方法具有很大的临床应用价值。然而食管是一种形态多变的器官,不同病人在不同时间都具有不同形态,而食管癌的位置和形态与周围组织和器官的关系密切相关。因此网络必须具有一定的空间定位能力、鲁棒性和泛化能力。此外影像数据本身存在噪声和伪影等问题,表明预处理工作的重要性不可忽视。这使得研究和开发一种准确、稳定、可靠的食管癌靶区自动勾画网络成为亟待解决的问题。本研究对食管癌靶区的自身特征进行了深入探究,提出了适合自动勾画食管癌靶区的网络。本文具体的研究内容和贡献如下:(1)本研究针对食管癌肿瘤边界模糊难以确定以及术前术后影像差异大等问题,提出了一种改进空间金字塔结构的食管癌肿瘤靶区自动勾画网络(SPPH-UNet)。该网络融合了食管癌肿瘤靶区的特征,并采用嵌入空间金字塔池化层模块的方法,实现不同尺度特征的融合,从而提高了对肿瘤边界的特征提取能力,更加准确地勾画出肿瘤靶区。同时,考虑到食管位置的变化和肿瘤解剖结构的复杂性,本研究采用了数据增强方法,以避免过拟合现象。实验结果显示,相较于经典的U-Net网络,SPPH-UNet网络在DSC和HD指标上分别提高了 8.2%和43.4%。(2)本研究针对二维网络无法联系部分与整体之间的空间依赖性问题,提出了一种嵌入胶囊模块的三维稠密网络(3D-DUCaps)用于自动勾画食管癌临床靶区。该网络在U-Net的编码层中嵌入胶囊模块,用来增强特征学习能力并保留更多信息,以便推算姿势和学习部分与整体之间的关系。此外引入稠密连接进一步促进高级语义信息和低级特征信息的融合,增强了网络的信息传递能力。相较于传统的二维深度学习网络,本文的三维深度学习网络具有更强的空间信息感知能力和更优秀的边界信息刻画能力,可以更好地勾画食管癌临床靶区。实验结果表明与经典的3D-UNet网络相比,3D-DUCaps网络在DSC指标上提高了 1.7%。更多还原

【Abstract】 In the field of medical image processing,automatic delineation of the target area for esophageal cancer is a highly studied problem.The determination of the target area has a direct impact on the effectiveness of radiotherapy and the incidence of toxic side effects.Traditional manual delineation methods require a lot of time and manpower and are susceptible to the subjective judgment and skill level of the doctors.Therefore,automated target area delineation methods have significant clinical value.However,the esophagus is a highly variable organ,and different patients have different shapes at different times,and the location and shape of esophageal cancer are closely related to surrounding tissues and organs.Therefore,the network must have a certain degree of spatial positioning ability,robustness,and generalization ability.In addition,there are issues such as noise and artifacts in the imaging data,indicating the importance of preprocessing work.This makes the development of an accurate,stable,and reliable automatic delineation network for esophageal cancer targets an urgent problem to be solved.This study conducted an in-depth exploration of the self-features of the esophageal cancer target area and proposed a network suitable for automatic delineation of the esophageal cancer target area.The specific research content and contributions of this thesis are as follows:(1)This study addresses the issues of fuzzy tumor boundaries and large pre-and post-operative imaging differences in esophageal cancer targets and proposes an improved spatial pyramid structure for automatic delineation of the esophageal cancer target area network(SPPH-UNet).The network integrates the features of the esophageal cancer target area and uses the embedding of the spatial pyramid pooling layer module to achieve the fusion of different scales of features,thereby improving the feature extraction ability for the tumor boundary and more accurately delineating the tumor target area.At the same time,considering the variation in esophageal position and the complexity of the tumor anatomy,this study adopts data augmentation methods to avoid overfitting.Experimental results show that compared with the classical U-Net network,the SPPH-UNet network improved the DSC and HD metrics by 8.2%and 43.4%,respectively.(2)This study addresses the problem of spatial dependency between parts and the whole that cannot be linked by two-dimensional networks,and proposes a three-dimensional dense network embedded with capsule modules(3D-DUCaps)for automatic delineation of the clinical target volume of esophageal cancer.The network embeds capsule modules in the encoding layer of U-Net to enhance the feature learning capability and retain more information for inferring pose and learning the relationship between parts and the whole.In addition,dense connections are introduced to further promote the fusion of high-level semantic information and low-level feature information,enhancing the network’s information transmission capability.Compared with traditional two-dimensional deep learning networks,the three-dimensional deep learning network proposed in this study has stronger spatial information perception ability and better boundary information characterization ability,and can better delineate the clinical target volume of esophageal cancer.Experimental results show that compared with the classic 3D-UNet network,the 3D-DUCaps network improves the DSC index by 1.7%.更多还原