基于卷积神经网络与非刚性匹配的骨盆通道特征提取及微创置钉研究

The treatment of pelvic fractures is the most difficult task in Orthopedics. The development of minimally invasive screw placement technology has provided a new approach for minimally invasive treatment of pelvic fractures. However, the corridor 3D reconstruction method is limited by the performance of traditional algorithm of the image segmentation, such as time-consuming, poor precision, and uncertain results. The corridor designing methods are subjectively influenced by doctors, and the results are quite different between each other. How to improve the efficiency and accuracy of methods of image segmentation and corridor establishment is the key to this technology. In our previous study, the deep learning techniques have been proposed and used for automatic segmentation of lower extremity, measurement of anatomical feature parameters, and the corresponding morphological study of bone and corridor. These put forward relevant solving ideas and methods of the morphology of pelvic corridor and the corresponding designing method. The present study intends to establish an automatic pelvic image segmentation and 3D reconstruction algorithm, which is completed through the application of the edge detection algorithm on the 2D image, combination and fusion of the 3D images with multiple size and features using the fully convolutional neural 3D U-net, practice and verification of the system using the manner of end-to-end method, and establish the statistical model using Procrustes analytical method, in order to improve the quality and speed of image segmentation. To establish the non-rigid 3D model matching algorithm, and establish the corresponding automatic algorithm to measure the pelvic corridors. To use the statistical average shape model based non-rigid 3D model matching algorithm to automatedly extract the anatomical features of 200 pelvic corridors for the anterior column of acetabulum, LC Ⅱ, and sacral-iliac joint screws, and establish the corresponding automatic algorithm to improve the efficiency of corridor designing. Establishment of the database of these pelvic corridors will theoretically and technically help to further measure and establish other databases of other corridors, such as Magic corridor, superior/inferior acetabular corridors, and other anatomical sites. To analyze the morphological parameters and summarize the subtypes of pelvic corridors, using unsupervised K-means algorithm, and establish the corresponding strategies for preoperative planning. To perform in vitro setup for verification of different subtypes. Based on these methods and results, we try to establish the pre-operative strategy of pelvic corridors designing for computer-assisted pelvic minimally invasive surgery, which will provide the important basis for future computer- or robot-aided surgery.

骨盆骨折治疗是创伤骨科最棘手的难题。通道螺钉置入技术的发展为骨盆骨折微创治疗提供了一项新手段。但目前螺钉置入通道的三维建模受限于现有图像分割算法耗时长、精度差、结果不确定等因素，通道规划则多受医生主观影响，结果差异大。如何提高图像分割和通道规划的效率和精度成为该项技术的关键所在。我们团队的前期工作利用深度学习技术完成了下肢骨性通道自动分割、特征参数测量与解剖亚型分析等解剖形态学研究，为进一步骨盆通道形态学研究及基于此的通道规划提供了解决思路。本研究拟利用结合边缘检测的全卷积神经网络进行骨盆自动三维重建，通过边缘检测算法对二维CT图像的应用，FCN 3D U-net对多尺寸特征的图像用skip-net进行融合，以及利用端对端的方式对系统进行训练和验证，进而提高图像分割速度和精度；利用Procrustes分析法建立骨盆统计学平均形状模型。构建非刚性三维模型匹配算法，建立通道解剖参数自动测量方法，提高通道规划效率；利用基于统计学平均形状模型的非刚性三维模型匹配算法测量200例骨盆的髋臼前柱、LC Ⅱ和骶髂关节通道的自身固有解剖特征参数，建立骨盆及三个骨盆通道的关键解剖特征参数数据库。为进一步测量和建立髋臼后柱Magic通道、髋臼上/下通道等其他骨盆微创通道或其他部位的骨性通道提供理论依据和技术支持。利用基于密度峰值的无监督K均值算法进行聚类分析，建立髋臼前柱、LC Ⅱ和骶髂关节通道亚型，建立髋臼前柱、LC Ⅱ和骶髂关节各亚型的微创螺钉置入术前规划策略，并选择不同亚型的骨盆模型进行体外置钉试验，验证各亚型微创通道的预测结果，据此初步制定计算机辅助骨盆微创通道螺钉置入术前规划。为计算机辅助骨盆微创手术的顺利实施提供重要依据。同时，为国人骨盆解剖特征库或其他部位骨骼三维模型数据库的建立提供理论基础，为今后机器人辅助骨盆微创手术提供科学依据。

骨盆骨折微创手术风险高、难度大，对手术规划和术中实施的精度要求极高。传统骨盆微创手术术前规划需在骨盆术前CT三维模型上进行，但现有图像分割与三维重建技术精度差，手术规划不同医生间差异大，术中闭合复位及微创固定的实施则过度依赖医生经验和个人技术。本项目针对骨盆微创手术中图像分割、手术规划、术中实施等关键环节所涉及的理论问题和技术难点进行研究，主要研究内容包括：1）通过基于局部和全局体积自注意力机制的新型卷积神经网络，自动分割与三维重建骨盆模型，利用Procrustes分析法建立202例骨盆统计学平均形状模型。2）基于骨盆统计学形态模型的配准网络自动建立骨盆骨折微创置钉通道。3）基于强化学习的骨骼通用配准框架进行骨盆微创手术复位规划并在股骨和胫骨中进行拓展应用。4）建立基于远程转心机构的计算机辅助骨盆骨折微创手术复位系统并完成实验验证。.项目的主要目标是围绕骨盆骨折微创手术中的图像处理和手术规划技术中的关键和难点问题的解决提供新理论、新方法和新技术，在骨盆CT图像自动精确分割、微创置钉路径自动建立、微创手术自动规划及精确实施等关键技术的研究上取得突破性进展。项目所涉及的基于点云的配准网络、基于强化学习的通用配准框架和其他关键性技术是计算机辅助骨科微创手术的共性问题和核心技术，具有重要的理论研究意义；所构建的基于远程转心机构的计算机辅助骨盆骨折微创手术系统，可进一步向自动化系统继续升级；所建立的核心算法和手术规划技术有待进行完善优化，并形成软件和硬件予以临床应用，从而发挥该研究的最大科学价值和社会效益。