基于深度学习的前列腺靶向穿刺活检中三维超声和MR图像实时精确配准方法研究

Biopsy is the gold standard for the definite diagnosis of prostate cancer.Three dimensional ultrasound and magnetic resonance (MR) fusion based targeted biopsy has become an important new technology for the diagnosis of prostate cancer. The accuracy of the targeted biopsy depends on the real-time and accurate non-rigid registration of multi-modal prostate images. Due to the complexity of multi-modal images and prostate deformation, the existing registration methods based on gray-level intensity, feature and model cannot resolve the above-mentioned image registration problem. This project aims at the study on the multi-modal non-rigid image registration method based on deep learning. The following research issues are involved. The deep learning model based on the nonlocal manifold learning network will be constructed to realize the collaborative structural representation of multi-modal prostate images,thereby transforming multi-modal image registration to mono-modal one. The construction of the statistical deformation model (SDM) of the prostate will be explored based on the multi-channel manifold learning network, thereby transforming the estimation of high-dimensional deformation fields to that of low-dimensional SDM coefficients. The semi-supervised manifold learning network will be constructed to represent the mapping relation between the collaborative structural representation results and the SDM coefficients, based on which the real-time and accurate deformation estimation for the multi-modal prostate images can be realized. The research work involved in the project will promote the advancement of the theory related to medical image processing and has important significance for improving prostate diagnosis accuracy and improving the well-being of the people. The research outcomes will find the wide application in such interventional surgeries as biopsy.

穿刺活检是确诊前列腺癌的金标准，三维超声和磁共振融合引导下靶向穿刺活检是前列腺癌诊断的重要新技术。精准靶向穿刺的关键在于前列腺多模图像的实时精确配准，但因多模图像和前列腺形变的复杂性，现有基于灰度、特征和模型的配准方法难以解决上述配准难题。本项目拟研究基于深度学习的非刚性图像配准方法，主要内容包括：构建基于非局部流形学习网络的深度学习模型，实现前列腺多模图像的协同结构表征，将多模图像配准演变为单模图像配准问题；探索基于多通道流形学习网络的前列腺统计形变模型构建方法，将高维形变场估计演变为低维形变模型系数估计问题；探索半监督流形学习网络构建方法，以挖掘协同结构表征结果和形变模型系数的映射关系，实现多模图像形变的实时精确估计。本项目的开展将有力推动医学图像处理相关理论的发展，对提高前列腺癌诊断精度，从而提升人民健康水平具有重要意义，研究成果在穿刺活检等介入式手术中具有广阔应用前景。

三维超声和磁共振融合引导下靶向穿刺活检是前列腺癌诊断的重要新技术，精准靶向穿刺的关键在于前列腺多模图像的实时精确配准。然而，因多模图像间存在结构对应性缺失、复杂的非线性灰度差异和非刚性形变，现有基于灰度、特征和模型的配准方法难以解决上述难题。针对前列腺多模图像间存在的非线性灰度差异，本项目提出了两类结构表征方法，包括基于中央凹的模态独立邻域描述子以及基于流形学习网络的数据自适应描述子，达到将多模图像演变为单模图像配准的目的。针对前列腺多模图像间存在的复杂非刚性形变，本项目提出了基于两级生成对抗网络（GAN）的多模图像监督配准方法，以及结合自由形变模型（FFD）和逆一致性约束的无监督图像配准方法。前者该在多通道GAN框架下，将多模图像配准问题演变成为结构表征估计和灰度估计两个子问题，以监督学习思想实现多模图像配准，避免了对形变场的直接估计，提升了配准效率和精度。后者通过FFD来降低配准网络的参数量，同时利用逆一致性约束实现双向配准，避免配准过程中图像拓扑结构的改变，可在无需形变场等金标准情况下，实现多模图像的高效无监督配准。针对前列腺多模图像间存在结构对应性缺失问题，提出了基于同时分割-配准的图像配准方法，该方法通过分割网络和配准网络分别实现前列腺组织提取及组织配准，对两个网络进行交替训练，最终达到同时分割配准的目的。基于仿真和临床数据，完成了图像配准算法性能评价工作，实验结果表明：本项目提出的基于深度学习的多模图像配准方法具有配准效率高，在目标配准误差、dice值等指标上优于传统基于互信息和结构表征的配准方法以及主流的基于深度学习的配准方法，展现出更优异的配准性能，为实现前列腺超声和MR图像的快速精确配准提供了有效手段。本项目的研究工作可有力推动医学图像处理相关理论的发展，对提高前列腺癌诊断精度，从而提升人民健康水平具有重要意义，研究成果在穿刺活检等介入式手术中具有广阔应用前景。