在MRI引导的放疗（MRIgRT）系统中，需要由MR图像数据预测对应的CT图像。建立不同模态图像之间的映射关系（模态映射），特别是MR-CT模态映射，具有重要的应用价值。当前的MR-CT模态映射方法主要利用少量同一病人的MR-CT图像数据对，通过图像分割、匹配和回归预测技术实现，尚不能生成与真实CT质量相当的伪CT图像。考虑临床可获取大量单模态图像，其中蕴含了图像模态的信息和知识，本项目提出采用深度学习方法，对大量单模态图像数据进行学习，建立提取图像深度表达的编码器和重构图像的生成器；通过少量MR-CT图像对，在不同模态图像的深度表达之间建立映射关系，以期生成高质量的伪CT图像，并对其应用于MRIgRT系统中放疗计划制定、剂量计算、在线引导等的有效性进行验证。本项目的研究还可为多模态图像配准提供一种新的解决途径，并对MR图像自动分析具有借鉴意义。

In MRI-guided radiotherapy (MRIgRT) systems, it is necessary to predict the corresponding CT images from MR image data. It is of great value to establish the mapping between different modalities of medical images (modal mapping), especially modality mapping between MR and CT. The current methods for MR-CT modality mapping mainly use image segmentation and regression prediction technology on the small dataset of MR-CT image data pairs of the same patient. These methods cannot generate pseudo CT images with the quality same as the real CT images. Considering the clinical access to a large number of MR and CT images, which contains the information and knowledge of the MR and CT image features and content, this project proposes to use deep learning method to learn the models for the deep image representation on the large dataset of single-modal image and the image generator which can reconstruct the images from deep image representations. Then, the mapping relationship is established between the deep representations of different modalities to achieve accurate modality mapping and to produce high-quality pseudo CT images on a small dataset of MR-CT pairs of the same patient. The effectiveness of the modality mapping model for radiotherapy in the MRIgRT system will be validated. The study of modality mapping between MR and CT can provide a new solution for multi-modal image registration. In addition, the study of deep representation of MR images can be helpful for MR image analysis.