Endoscopic navigation using ORB-SLAM for urethral minimally-invasive surgery (ESP: Navegacion endoscopica con ORB-SLAM para cirugia uretral minimamente invasiva)

Kurzfassung

In minimally invasive surgery, an endoscope is moved through a body cavity or organ. The aim of SLAM (Simultaneous Localization And Mapping) is to estimate the 3D map of the cavity or organ being explored and to locate the endoscope with respect to the map. The only information used for this purpose is the video taken by the endoscope. ORB-SLAM is a SLAM system developed at the University of Zaragoza that uses FAST features and ORB descriptors to estimate both the 3D map and the camera pose. The main problem of ORBSLAM2 is that this system is robust (with high accuracy and without failures or deviations) when it works with scenes that can be assumed rigid, with texture, in which it is easy to extract and match visual features. This is not the case with organs, as they are deformable, wet and it is difficult to visually find reliable features (corners) within them, which can lead to an incorrect map estimation. Abrupt changes in illumination and sudden movements of the endoscope must also be taken into account. In this work, ORB-SLAM2 has been modified to work in sequences recorded by an endoscope. This modifications include reducing the number of points required to perform the monitoring, creating a more robust initial map and pre-processing the images to increase contrast, reduce noise and avoid detecting points in the reflections. In addition, to achieve more robust keypoints, different detectordescriptor combinations have been tested, creating a new vocabulary for each of them. It has been demonstrated that with the modifications made to ORB-SLAM2, the change from FAST-ORB to AKAZE-ORB and the new vocabulary, it has been possible to estimate 80 % of the trajectory when with the original system it was not possible to estimate 40 %. This improvement has only increased the computation time by 40 ms. A SLAM system can be very useful in surgical interfaces since it allows to connect a point in screen with the estimated 3D map, allowing the surgeon to insert augmented reality annotations in the appropriate places.