Automatic detection of surgical gauzes using Computer Vision

In this paper we study the effectiveness of different algorithms for texture classification based on the Local Binary Patterns method, in order to obtain in future a tracking system for surgical gauzes during a laparoscopic operation. Due to the mobility of the camera and the lack of a precise control over its position, the algorithms must work under unknown illumination and viewpoint parameters. Our intention is to provide to surgeons a simple tracking system in order to avoid mislaid gauzes during the operations, which might be completely unattended so that the surgical team can concentrate on the patient and not on the gauze counting. Due to blood stains, color classification it is not possible, also we cannot use the shape of the gauze, neither its position relative to the camera nor its size. So, the only possibility is to obtain a texture classification algorithm able to discriminate between the gauze surface and the background of the scene, i. e. the interior of the patient, laparoscopic tools and whatever is not a surgical gauze. For this purpose we try a few operators, looking for an algorithm for gray-scale texture classification. We apply the rotation invariant of the well-known Local Binary Pattern (LBP^riu2), an improved version (NI-LBP^riu2) and a control operator consisting of a direct comparison between the histogram of a region of interest and the histogram of a reference image, which we call GauzeTrack Local Histograms Algorithm. All mentioned algorithms have been applied both on test textures and images extracted from a video of a real laparoscopic surgery. This way we could ensure that the proposed method works on a real situation and not only in the laboratory.