基于机器学习和几何变换的实时2D/3D脊椎配准

在图像引导的脊柱手术中，实时高效的2D/3D配准是一项重要且具有挑战性的任务.通常的2D/3D配准一般是将三维图像投影到二维平面，然后进行2D/2D的配准.由于投影空间涉及到3个平移以及3个旋转参数，其投影空间的复杂度为O（n6），使得配准很难兼具高准确性和高实时性.本文提出了一个结合机器学习与几何变换的2D/3D配准方法，首先，使用统计形状模型对目标脊椎进行建模，并构建了一种新的投影方式，使得6个投影参数中的4个可以使用几何的方法计算出来；接下来利用回归学习的方法学习目标脊椎的形状与投影参数之间的关系；最终，结合学到的关系和几何变换完成配准.本方法的两个姿态参数的平均预测误差为0.84°和0.81°，平均目标配准误差（Mean target registration error，mTRE）为0.87mm，平均配准时间为0.9s.实验结果表明本方法具有很好的实时性和准确性.

Real-time 2D/3D Registration of Vertebra via Machine Learning and Geometric Transformation

In spine operations, an effective 2D/3D registration is of great importance yet a challenging task. Traditional 2D/3D registration approach projects 3D data to a 2D plan then conducts a 2D/2D registration. It is difficult to obtain both high accuracy and good real-time performance, because the projection space has 3 translation and 3 rotation degrees of freedom and its complexity O(n6). In this paper we propose a method which combines machine learning strategy with geometric transformation. We build a shape model using statistical shape model and construct a new projection method, which makes it possible to calculate that 4 of the 6 projection parameters with the geometric method. Then we use regression learning method to learn a pose model between shape of vertebra and projection parameters. Finally we fulfill the registration by using the learned pose model and geometric transformation. Using the proposed method, the mean predicted errors of two pose parameters are 0.84° and 0.81°, respectively. The mean target registration error (mTRE) is 0.87mm. And the mean time of registration is 0.9s. These results show that our method owns both high accuracy and good real-time performance.