Adapting monocular SLAM for surface-trace free 3D registration in navigated arthroscopy

Video-based surgical navigation (VBSN) relies on registering a preoperative 3D model to intraoperative data to overlay the surgical plan onto the patient's anatomy, providing real-time guidance. In current systems, registration is achieved by tracking two visual markers: one rigidly attached to the anatomy, called the anatomy marker, and another mounted on a digitizing tool. The surgeon manually collects 3D points by contacting the anatomical surface with the tool tip. This workflow requires anatomy and tool markers to be simultaneously visible and continuous tool-tissue contact, which is time-consuming and may pose risks to the patient. In an attempt to overcome these difficulties, this paper investigates the possibility of removing visual markers from VBSN.

Recent monocular Simultaneous Localization and Mapping (SLAM) approaches are adapted to the arthroscopic context to estimate both the camera pose and a 3D reconstruction of the scene from arthroscopic video.

Experimental results indicate that the anatomy marker cannot be removed, as camera pose estimation using SLAM does not simultaneously achieve the accuracy and real-time performance required for surgical navigation. However, the digitizing tool can be eliminated, enabling contactless registration using a SLAM-based variant. Moreover, SLAM can recover both structure and motion in short time segments without the anatomy marker, allowing the surgeon to inspect regions of the anatomy without constantly maintaining the anatomy marker in the field of view.

The proposed SLAM-based approach advances VBSN by enabling real-time, contactless registration compatible with clinical requirements, as validated in cadaveric arthroscopy experiments with registration accomplished on the fly.