Localization of flexible surgical instruments inendoscopic images using machine learning methods

Abstract: The work presented in this document deals with Machine learning algorithms used in surgical robotic problems, especially here in the Isis project. The aim of the project is to replace manual handles for endoscopic operations by a set of motors commanded via high-tech interface. The main aim of the thesis is to solve the problem of the estimation of the pose of flexible surgical instrument, using only the video flux from an endoscopic camera. After a short introduction about the system its environment and the definition of the pose problem, the work is divided in two chapters. Machine learning algorithms and learning systems are used in both parts. The first chapter deals with image processing and video tracking. Usage of colored markers and how the learning is made to perform the best segmentation is explained. It starts with simple linear classification to end with an Adaboost algorithm. The learning database construction and all the challenges it raises are explained too. Then the tracker used in the system is decomposed and its structure explained. The results of the whole tracking system are presented at the end of the chapter. The second chapter deals with function approximation: we build Radial Basis Function networks in order to approximate the final position of the instrument (or its mechanic parameters) from the data extracted by the tracking system. Learning algorithms are used too. The learning training set is built in a laboratory environment where the real position of the instrument can be measured. The last chapter is a collection of improvements that could be added to the system and opens on future perspectives about the project in general.