面向早期肝癌立体定向放疗的多模态图像处理关键技术研究

Accurate localization of liver tumor in abdominal CT and cone beam CT (CBCT), and adaptive tumor tracking during radiotherapy are essential to stereotactic body radiation therapy planning and high-precision treatment for patients with early hepatocellular carcinoma. Aiming at the inseparable of liver tumor from CT and CBCT sequences, the project proposes an abdominal MRI sequences guided automatic sketching method. Firstly, liver tumor is segmented accurately from MRI sequences by combining the multi-scale based 3D Convolutional Neural Networks (3D-CNN) and fully connected Conditional Random Field (CRF). Then, the coarse registration of multi-modality images is conducted according to the structure of liver vessels. Next, the Markov Random Field (MRF) modeled fine registration framework is optimized by the Graph-Cut algorithm to acquire accurate liver tumor contours in CT and CBCT images. Aiming at the inaccurate localization of target in radiation therapy because of the respiratory motion and gastrointestinal motility, the project proposes an adaptive liver tumor tracking method. Firstly, a multi-resolution B-spline deformable image registration algorithm is applied to construct the original respiratory motion model. Next, to track the tumor adaptively, the baseline and deformable vector fields (DVFs) of the motion model are updated respectively according to the inter-fraction and intra-fraction tumor motion. This project will not only help the radiologists locate liver tumor target accurately and therefore formulate radiotherapy planning reasonably, but also contribute to improving the precision of radiotherapy, preventing normal tissue and organs from being exposed to particle beams, and thus, improving overall survival and quality of life for patients with early hepatocellular carcinoma.

对腹部CT和锥束CT(CBCT)图像中肝脏肿瘤的精确定位，以及放疗时肿瘤的自适应追踪是早期肝癌立体定向放疗计划制定和精准治疗的关键。针对CT和CBCT图像中肝脏肿瘤难以精确分割的问题，提出腹部MRI图像引导的CT和CBCT图像肿瘤自动勾画方法。首先基于多尺度3D-CNN和全连接CRF实现MRI图像肿瘤的准确分割，然后根据肝脏主血管结构实现腹部多模态图像粗配准，最后研究基于图割求解的MRF精配准框架，准确勾画CT和CBCT图像中肝肿瘤区域。针对放疗时因呼吸运动、胃肠蠕动等导致的靶区定位不准，提出肝脏肿瘤自适应追踪方法。首先基于多分辨率B样条形变配准构建初始呼吸运动模型，然后根据肿瘤在分次间和分次内的运动情况对该模型进行基线和形变向量场更新，实现对肿瘤的自适应追踪。该研究将辅助医生精确定位肝癌放疗靶区并制定放疗计划，提高放疗精度，减少射线对正常组织和器官的损伤，提升患者的生存率，改善生活质量。

对腹部CT和CBCT图像中肝脏、周边器官、肿瘤的准确分割，以及放疗时肿瘤的精准定位是早期肝癌立体定向放疗计划制定和精准放疗的关键。.本项目针对CT和CBCT图像中肝脏、肝脏周边器官由于边界模糊难以精确分割的问题，提出了基于水平集和形状描述符的腹部CT序列肝脏自动分割，基于水平集和稀疏形状组合的腹部CT序列肝脏分割，基于方向信息的腹部CT图像肝脏区域自动分割，以及基于图划分与三维Chan-Vese模型的腹部CT序列图像多器官自动分割等方法。.早期肝癌立体定向放疗需要对腹部不同模态的图像(如MRI、CBCT、CT)进行配准以达到信息融合。考虑到个体间及个体内的腹部软组织随姿势变化和呼吸运动会发生较大形变，腹部CT图像目标器官的配准面临巨大挑战，提出了基于深度学习的腹部CT图像目标器官配准，基于全卷积网络的三维腹部CT图像多器官配准，基于无监督学习的腹部CT图像多器官配准，以及基于B样条和Laplacian算子结合的CT/MR图像配准等方法。.针对因成像噪声以及CT/MRI图像中肝脏肿瘤边界模糊、结构复杂、灰度多样等给肿瘤区域勾画带来困难，以及放疗时因呼吸运动、胃肠蠕动等导致的靶区定位不准，提出了不同的肝脏肿瘤分割与定位方法：基于LI-SLIC超像素的腹部图像肝脏肿瘤自动分割，基于置信连接区域生长和图割的肝脏肿瘤半自动分割，腹部CT序列图像肝脏肿瘤自动分割及立体定位，以及基于深度学习和注意力机制的三维医学图像肝脏肿瘤立体定位及跟踪等方法。.该研究成果可辅助医生及时有效地获取肝脏及肿瘤的整体信息和三维显示，准确定位放疗靶区并制定放疗计划，提高放疗精度，减少射线对正常组织和器官的损伤，为早期肝癌立体定向放疗提供技术支持和决策服务，从而提升患者的生存率，改善生活质量。.本项目已发表学术论文17篇，其中SCI检索12篇，EI检索中文权威期刊4篇；申请国家发明专利12项，授权8项；软著2项；培养博士后1人，毕业研究生12人，其中博士1人。