Qualitative Comparison of Image Stitching Algorithms for Multi-Camera Systems in Laparoscopy

Multi-camera systems were recently introduced into laparoscopy to increase the narrow field of view of the surgeon. The video streams are stitched together to create a panorama that is easier for the surgeon to comprehend. Multi-camera prototypes for laparoscopy use quite basic algorithms and have only been evaluated on simple laparoscopic scenarios. The more recent state-of-the-art algorithms, mainly designed for the smartphone industry, have not yet been evaluated in laparoscopic conditions. We developed a simulated environment to generate a dataset of multi-view images displaying a wide range of laparoscopic situations, which is adaptable to any multi-camera system. We evaluated classical and state-of-the-art image stitching techniques used in non-medical applications on this dataset, including one unsupervised deep learning approach. We show that classical techniques that use global homography fail to provide a clinically satisfactory rendering and that even the most recent techniques, despite providing high quality panorama images in non-medical situations, may suffer from poor alignment or severe distortions in simulated laparoscopic scenarios. We highlight the main advantages and flaws of each algorithm within a laparoscopic context, identify the main remaining challenges that are specific to laparoscopy, and propose methods to improve these approaches. We provide public access to the simulated environment and dataset.