Keypoint Matching for Instrument-Free 3D Registration in Video-Based Surgical Navigation

Tânia Baptista, Carolina Raposo, Miguel Marques, Michel Antunes, João Pedro Barreto

Abstract: Video-based Surgical Navigation (VBSN) inside the articular joint using an arthroscopic camera has proven to have important clinical benefits in arthroscopy. It works by referencing the anatomy and instruments with respect to the system of coordinates of a fiducial marker that is rigidly attached to the bone. In order to overlay surgical plans on the anatomy, VBSN performs registration of a pre-operative model with intra-operative data, which is acquired by means of an instrumented touch probe for surface reconstruction. The downside is that this procedure is typically time-consuming and may cause iatrogenic damage to the anatomy. Performing anatomy reconstruction by using solely the arthroscopic video overcomes these problems but raises new ones, namely the difficulty in accomplishing keypoint detection and matching in bone and cartilage regions that are often very low textured. This paper presents a thorough analysis of the performance of classical and learning-based approaches for keypoint matching in arthroscopic images acquired in the knee joint. It is demonstrated that by employing learning-based methods in such imagery, it becomes possible, for the first time, to perform registration in the context of VBSN without the aid of any instruments, i.e., in an instrument-free manner.