食管癌调强放疗靶区勾画中CT与DWI图像非刚体配准方法研究

The delineation accuracy of gross tumor volume (GTV) determines whether the expected therapeutic effect of Intensity Modulated Radiation Therapy for esophageal cancer can be achieved, and image registration is the key factor to achieve the accurate delineation of GTV. Due to the complex pathological features such as mucosal folds, it is difficult to achieve effective registration of esophageal cancer images using the existing methods. This proposal intends to leverage the latest developments in deep learning for the clinical demands, to research the intelligent registration method of CT and DWI images of esophageal cancer: (1) Considering the insufficiency of clinical data of esophageal cancer and the modalities discrepancy, the generative models based on generative adversarial networks and modality style will be constructed to augment the dataset of esophageal cancer; (2) Concerning the poor universality and inaccuracy of the handcrafted features, an algorithm of feature representation based on 3D U-shaped residual networks will be presented to extract the deep features of images for esophageal cancer; (3) In view of the extensibility of Shannon entropy and the non-smooth deformation fields in the images of esophageal cancer, a similarity measure based on generalized entropy and a spatial regularization based on the model of esophageal mucosal folds will be constructed, to address the problems of the similarity measure of different modalities and the non-smooth deformations in images. Finally, a non-rigid registration framework based on deep learning and the similarity measure of generalized entropy will be built to achieve the accurate registration for CT and DWI images of esophageal cancer. This project is expected to promote the intelligence of clinical diagnosis and treatment, and advance the theories and applications of medical image processing.

GTV勾画精度决定食管癌调强放疗能否达到预期治疗效果，其中图像配准是GTV准确勾画的关键。由于包含黏膜皱襞等复杂的病理特征，造成现有方法难以实现食管癌影像的有效配准。本项目拟将深度学习领域的最新进展与临床需求相结合，深入研究食管癌CT与DWI影像的智能配准方法：(1)针对食管癌临床数据的不足及模态差异，拟构建基于GAN和模态迁移的图像生成模型，扩充食管癌影像数据集；(2)针对人工设计特征普适性差、准确性低的特点，拟提出基于三维U形深度残差网络的特征表征算法，提取食管癌影像的深层特征；(3)针对香农熵的延展性及食管癌图像的非平滑形变场，构造广义熵相似度和基于食管黏膜皱襞模型的空间正则化，以解决多模态相似度与非平滑形变问题。最终构建基于深度学习和广义熵相似度的非刚体配准框架，实现食管癌CT与DWI图像的精确配准。相关研究成果有望提升临床诊疗的智能化水平，并对医学图像处理理论与应用起到推动作用。

本项目将深度学习领域的最新进展与临床需求相结合，深入研究多模态食管癌影像的智能配准与融合方法，在相关研究任务上主要完成了如下工作。1）针对人工设计特征普适性差、准确性低的特点，提出了基于深度密集卷积网络的特征表征算法，提取多模态医学影像的深层特征；2）针对深度神经网络中空间和通道信息丢失的问题，构建了基于双注意力机制的多模态医学图像融合模型；构造了三叉戟膨胀感知模块，精确地确定特征的位置信息，提高了网络的特征表示能力，此外，为了充分利用中间层信息，构建了一个双残差超密集模块，结合超密集压缩分解网络提出了一种新的端到端图像融合模型，以实现不同路径之间的信息交互；提出结合轴向注意力的特征提取网络，该网络可以在提取多尺度特征的同时捕获长程语义信息。此外，提出了一种新的融合策略，构建了熵注意力，将熵特征应用于边缘和细节特征聚合，在融合过程中保留图像的完整目标，缓解信息丢失的问题，使图像的边缘细节更丰富；3）在其他相关任务上的工作：提出了一种基于边界监督和多尺度特征聚合的医学图像分割网络，构建边界监督模块增强病灶与正常组织的低对比度，此外，多尺度特征融合用于提高不同尺度病灶的分割能力，最后采用上下文聚合模块整合高层和低层的图像特征，并生成全局信息。上述研究为后续食管癌靶区的智能自动勾画、准确定位提供有效支持。