A Comparison of Surgical Cavity 3D Reconstruction Methods

Abstract

In robot-assisted minimally invasive surgery (RMIS), surgeons are tasked with navigating and operating in dynamic 3D cavities. 3D spatial information can specifically indicate resection targets or delicate structures to avoid, and is critical to many surgical tasks. While preoperative imaging can provide information about relative anatomical or spatial locations of interest, these data are only a snapshot of a dynamically changing environment; natural movement due to breathing, deformation or natural heartbeat require the operator to make intraoperative, on-the-fly decisions. Online monocular RGB visual feedback may assist, but lacks the spatial information critical to precise surgical operations. In reality, surgeons must negotiate static, offline preoperative data with reduced dimensionality or sparse imaging to perform intricate, life-critical tasks. Real-time dense 3D reconstruction can alleviate these issues. In this comparative work, three approaches towards dense 3D reconstruction from laparoscopic imaging are investigated. For this, a new laparoscopic surgical scene dataset is introduced, and was captured using a pre-calibrated stereo camera. This camera recorded a surgical scene from various viewpoints in time, and a ground truth dense surgical scene was recorded with a high-precision 3D scanner. Three different reconstruction algorithms were then implemented and compared, including simultaneous localization and mapping (SLAM), visual odometry (VO), and structure from motion (SFM). A successful method will enable real-time registration of preoperative imaging and segmentation, thus resulting in safer and more robust surgical operations. This work also enables future research in vision-based force estimation and other research areas in RMIS.