Comparatively Evaluating Stereo-Matching Algorithms for Minimally Invasive Surgery

3D reconstructing surgical scenes for medical robotics is a desired capability to allow new visualization methods through mixed reality and enable a higher level of autonomy. A popular approach to achieving 3D reconstruction is stereo-matching with binocular cameras. Over the years, many different approaches to stereo-matching have been proposed for non-surgical robotic tasks. Their performance in Minimally Invasive Surgery (MIS) tasks, which is often more challenging due to various image artifacts, has not been sufficiently studied. As such, this paper evaluates the performance of various stereo-matching algorithms, including traditional and deep learning-based approaches, on stereo-endoscopic datasets (quantitatively on SERV-CT and SCARED; qualitatively on Hamlyn). It was found that performance between algorithms varied significantly, and the lack of a large labeled stereo-endoscopic dataset severely affected the performance of some deep learning models. Image artifacts such as texture-less surfaces, specularities, smoke, motion blur, and other obscuring features remain difficult challenges for stereo-matching algorithms. The results of this study could provide directions for further development of stereo-matching algorithms for MIS and encourage the creation of more stereo-endoscopic datasets.