基于深度时空特征融合的呼吸运动建模方法研究

The respiratory motion model plays an important role in the image guided radiation therapy (IGRT) of lung cancer. Its modeling is accomplished by registering a spatial-temporal image of a complete breathing cycle to a unified coordinate system. Lung motion is highly nonrigid. Moreover, lung imaging has the following characteristics: low contrast, motion artifact and high noise level. This makes the modeling of spatial-temporal nonrigid deformation under multiple interference a very challenging problem. This project aims to build an accurate respiratory motion model, the main research contents and innovations include (1) image registration based on unsupervised deep learning with prior knowledge, which can improve the accuracy of the registration by improving the existing algorithms that can not reflect the constraints and conditions of the actual application. (2) feature representation and fusion based on space-time subspace, which uses multi reference images to establish spatial-temporal subspace and spatial-temporal modeling, and improves the accuracy of modeling by improving the existing algorithms that relies on average image or single reference image without spatial-temporal modeling. (3) modeling of respiratory motion by fusion of deep spatial-temporal features and prior knowledge, which further optimizes of the modeling results with a priori to build an accurate respiratory motion model. The modeling method proposed in this project is an important supplement and expansion of the existing modeling methods of respiratory motion, which has important theoretical research and practical application value.

呼吸运动模型在肺癌的图像引导放射治疗(IGRT)中具有重要的作用。其建模主要通过将一个完整呼吸周期的时空图像配准到统一坐标系来完成。肺部运动具有高度非刚性的特点，而且肺部图像存在对比度低、运动伪影和噪声水平高等干扰，这使多干扰情况下准确建模时空非刚性变形成为一个非常具有挑战性的问题。本项目以构建准确的呼吸运动模型为目标，主要研究内容和创新点包括 （1）融合先验的无监督深度学习型图像配准，改进现有算法无法体现实际应用的约束和条件的不足，提高配准的准确性；（2）基于时空子空间的特征表示与融合，利用多参考图像建立时空子空间，进行时空建模，改进现有算法依赖于平均图像或单一参考图像而没有时空建模的不足，提高建模的准确性；（3）融合深度时空特征和先验的呼吸运动建模，融合先验进一步优化建模结果，从而构建准确的呼吸运动模型。本项目提出的建模方法是现有方法的重要补充和扩展，具有重要的理论研究与实际应用价值。

呼吸运动建模被用来描述肺部的运动情形，主流方法是时空图像配准。针对瓶颈问题，本项目完成了：.（1）无监督学习框架的研究 提出了一种基于卷积神经网络（CNN）的图像配准方法和一种基于视觉自注意力模型 (ViT) 的图像配准方法, 实验验证了这两种配准方法的准确性。.（2）引入先验知识的研究 针对当前研究忽略利用历史知识的问题，提出了一种层次化的迁移学习方法来引入历史知识；针对忽略利用人工标注的问题，提出了一种基于像素加权的先验知识引入方法。实验验证了两种引入先验知识方法的有效性。.（3）原型系统的研究 基于图像配准的结果， 我们进行了呼吸运动建模的研究，构造了一个完整的原型系统。利用该系统，我们实现了输入一张胸部时空图像, 输出呼吸运动模型的研究目的。.本项目的创新之处在于构建了基于无监督学习的配准框架和在医学图像处理中引入先验知识，为相关的学术研究提供了新思路。本项目的成果将进一步应用到医学图像处理领域，面向人民身体健康，促进智慧医疗的发展。