基于超限学习机2.5D-3D配准的脊椎手术导航

摘要

基于视觉的自动导航是自动手术系统和手术机器人的关键组成部分.手术前基于CT扫描重建病人组织器官的三维模型,术中采集对应器官的二维影像,使其与三维重建模型进行实时匹配和对准,即实现虚-实配准,即可实现计算机视觉辅助的手术导航.面向脊椎这一刚性组织,研究观察影像与三维重建模型的自动刚性配准.为实现快速、鲁棒的特征提取和匹配,提出使用超限学习机构建的深度学习模型,基于大量训练数据学习2.5D深度图像和3D模型关键点的特征提取,并进行三维配准.人体脊椎数据库的配准实验和猪脊椎模拟手术导航的结果表明,上述方法具有训练时间短、匹配精度高、运行速度快等特点,非常适合于脊椎手术导航.%Vision-based automatic surgery navigation is a key component of autonomous surgery systems or surgery robots.The surgeon performs CT scanning and 3D reconstruction for the pathological tissue of the patient before surgery.During surgery,2D images of the corresponding tissue are captured,which are matched and aligned against the reconstructed tissue in real-time.Such virtual-real alignment is a critical step for visual surgery navigation.This paper studies the virtual-real alignment problem for the rigid tissue of human spines and geometry-based alignment against the reconstructed 3D model.The Extreme Learning Machine (ELM) based deep networks is used for 3D critical point extraction and feature computation,based on a large amount of training data.We demonstrate the efficiency and effectiveness of our method in registration experiment of human spine database and automatic navigation on simulated pig spine surgery.