Reconstruction de courbe D à partir d une vue

Abstract

Information processing is used to deal with various problems that exist in many fields in modern science. Among these areas, there is the field of medical engineering, in which we look for take advantage of technology to improve medical tools, so as to ease the work of health professionals. Then there is the field of computer vision, a branch of artificial intelligence and computer automation that looks for developping software that can capture and understand the information contained in the images. During the internship, I have embarked on a research about these two fields, with the objective of collaborating in a project that brings together these two themes. The project focuses on developing a solution to the problem presented when tracking a catheter that goes through the vascular system, using x-ray images that are taken in real time. This problem is translated to a D monocular reconstruction problem, which is about finding the shape of the catheter curve in space according to the information captured in a single direction of the camera. The final purpose of this project is to provide the software that allows the augmented reality display of the catheter, so as to be useful for doctors during a medical intervention. To touch on this problem, we have opted for the use of Bayesian methods, which allow the mechanical and physical aspect of the studied system to be treated separately of the integration aspect which uses the information subtracted from the radiographic images. Thanks to the work of several researchers who have participated in this project, at the beginning of the internship there was already some important advance on the first aspect. About the second aspect, there was not a full dedication for this aspect until that moment. In this internship, I dedicated myself to the study of this second aspect, with the objective of providing methods to deal with this part of the project. With the aim of generalizing the existing types of filtering, I proposed a generic formal method, which allows to carry out the vupdate step with the information of the radiography images by minimizing any cost function. The difficulty in the monocular reconstruction is mainly in the fact of using a single camera, which will be static, so there will be an orthogonal direction in which it will be difficult to capture the information of the position in this direction. This type of problem is affordable by making assumptions about both the shape of the catheter and its mechanics. The use of some point of which position is known is also a way to deal with this problem. During the internship, a first step of research allowed me to acquire the knowledge on the already existing methods, so as to advance on the solution to our problem. I worked in the implementation of different methods of Bayesian filtering, to the point of being able to collaborate by providing software that is used to implement these methods. More precisely, it is about a battery of filtering methods that can be used for the reconstruction of the catheter curve. The study and adaptation of these methods led me to the design and implementation of the CSF method, which arose from the need to count with an option that works directly with radiographic images, without having to perform previously segmentation of the catheter. Details on the design, implementation and testing of this method are presented in this report.