面向图像引导放射治疗的腹部器官自动分割方法研究

In recent years, the incidence and mortality of intra-abdominal organ tumor have increased year by year. Radiotherapy is one of the major scheme for malignant treatment. Accurately delineating the boundaries of intra-abdominal organs can help to evaluate the tolerated dose for abdominal organs and reduce radiation toxicity, which has an important clinical diagnostic significance. In the current image-guided radiotherapy planning system, physicians delineate the multi-organ contours by hand, which is time-consuming and can cause sketching errors due to their difference experience. And the delineating performance by hand cannot be repeated and affected by subjective factor. In order to solve these problems, we propose an automatic abdominal multi-organ segmentation for image-guided radiotherapy in the segmentation framework using the full convolutional network (FCN). It includes as follows: (1) Deep boundary-learning model for automatic abdominal multi-organ segmentation is proposed. Firstly, the sequential learning FCN is proposed to solve the problem that the FCN model don’t deal with sequence coherence for image segmentation. Secondly, we propose the deep boundary-learning convolution neural network to solve segmentation boundary inaccuracy. Thirdly, the semantic segmentation model based on conditional random field is establish to obtain the multi-organ space information. (2) In order to the train the small labeled datasets, we propose the semi-supervised automatic abdominal multi-organ segmentation by introducing the generative adversarial nets.

近年来腹腔内器官肿瘤发病率和死亡率逐年增长，放疗治疗是恶性肿瘤主要治疗手段之一。准确勾画出腹腔内器官的边界，有助于准确评价各腹部器官的受剂量，降低放疗毒副反应，具有重要的临床诊断意义。在当前图像引导放射治疗计划系统中，医师手工勾画腹部器官存在耗时、经验差异引起的勾画误差、不可重复且受医生主观性影响较大等问题。为解决这些问题，本项目拟在全卷积神经网络分割框架基础上，提出面向图像引导放射治疗的腹部器官自动分割方法。具体研究内容如下：(1)研究时序学习的全卷积神经网络分割、深度边界学习的卷积神经网络和基于条件随机场的语义分割模型，建立腹部器官自动分割的深度边界学习模型，解决卷积神经网络分割模型无法解决序列图像相干性、分割边界不精确和像素空间信息丢失问题；(2)在腹部器官自动分割的深度边界学习模型基础上，引入半监督学习生成式对抗网络，进一步提出半监督腹部器官自动分割模型，解决小样本训练问题。

为准确评价各腹部器官的受剂量，降低放疗毒副反应，解决医师手工勾画腹部器官存在耗时、经验差异引起的勾画误差、不可重复且受医生主观性影响较大等问题，本项目拟在主动轮廓模型和全卷积神经网络分割框架基础上，提出半自动分割方法和全自动分割方法，研究要点如下：(1)为解决深度学习方法需要大量标注训练样本，研究了基于模糊区域性主动轮廓模型和局部混合型主动轮廓模型分割方法，该成果发表在期刊Information Science、Applied Soft Computing等期刊上；(2)研究了基于注意力机制的自动分割模型，其研究成果发表在Computers in Biology and Medicine等期刊上；（3）为获取图像像素间长距离依赖关系，提出了结合Swin Transformer和MLP等网络的优势，已形成科研成果将发表在国际重要期刊上。通过项目的实施，共发表论文18篇；培养研究生13名，完成5篇硕士论文。