将二维血管造影DSA图像与三维CT/MRI图像高精度配准融合，既可获得清晰且包含细小血管的血流动态信息，又可得到血管及周围组织的三维结构，一直是血管外科精准诊疗的目标，但现有方法尚难以满足临床需求。本项目旨在围绕该领域存在的挑战性问题，探索通过深度学习方法的创新和应用实现该目标的新途径。针对医疗影像分割和配准领域标注数据缺乏以及深度学习方法的弱归纳偏置等问题，在现有的迁移学习、强化学习、生成对抗网络、图神经网络、元学习等深度学习方法上突破和创新，提出基于强化学习和网络自学习先验的高稳定去运动伪影算法、基于多源数据迁移的单视角DSA图像主干血管三维重建、基于知识推理和注意力机制的半监督血管分割及修正、基于多智能体强化学习的非刚性精细配准和融合以及基于深度先验和元学习的医学影像自动分割等，进而实现一个基于深度学习的DSA与CT/MRI配准融合及三维组织分割系统，为血管外科医生精准诊疗提供支撑。

It has been the goal of precision vascular surgery to obtain the clear blood flow information, the three-dimensional structure of the blood vessels and the peripheral tissues by the high-precision registration and fusion of the two-dimensional digital subtraction angiography (DSA) images and the three-dimensional enhanced CT/MRI images. However, the existing methods are still difficult to meet clinical requirements. The project aims at exploring new approaches in the deep learning field to provide solutions for the challenging topic. First, we propose a robust motion deblurring algorithm based on reinforcement learning and self-learning priors to improve the quality of image segmentation and registration. Second, for lacking labeled data in the medical registration and segmentation fields, we propose a novel method for 3D main vessel reconstruction based on DSA images by multi-source data transfer learning. Third, a semi-supervised vessel segmentation and error correction method is proposed based on knowledge reasoning and attention mechanism. Fourth, the non-rigid registration and fusion algorithm based on multi-agent reinforcement learning is explored to achieve high-precision registration of DSA and CT/MRI images. Finally, in view of the weak inductive bias in the existing deep learning methods, we present the automatic segmentation of CT/MRI medical images based on deep prior and meta-learning. The project aims to make breakthroughs in deep learning methods such as transfer learning, reinforcement learning, graph neural networks, and meta learning and finally develops a deep learning-based DSA and CT/MRI registration and three-dimensional tissue segmentation system to support the precise diagnosis and treatment for vascular surgeons.