Autocalibrage de sequences d'images thoracoscopiques

The considerable expansion of video imaging techniques in recent years has opened the way toward exploring new surgical techniques in the medical field including minimally invasive surgery approaches.;3D navigation systems are currently being developed to assist the surgeon in the fields of cardiology and neurosurgery. However, most 3D navigation systems for spine surgery are designed to assist conventional surgeries requiring large openings. In addition, these systems require a set of markers attached to the vertebrae to register a pre-operative 3D model with peroperative images. The procedure for setting the markers is not only invasive and thus complicates the surgical protocol but cannot be performed in a minimally invasive surgery. Therefore, the ultimate goal of our project is to provide a framework for assisting minimally invasive surgeries of the spine, based solely on information contained in the sequence of peroperative images acquired by an endoscopic camera inserted into the patient's chest through a small incision.;Our goal is thus to develop a self-autocalibration method for a sequence of thoracoscopic images acquired during a minimally invasive surgery of the spine. This will allow us to readjust a 3D pre-operative patient spine model to the images of the sequence in real time as the surgeon performs the operation. The thoracoscopic images will be repositioned in 3D space to allow the surgeon to move around the 3D model and thus facilitate the operation. He will be able to know how close his surgical instruments are to critical structures (e.g. the aorta, the vena cava or the thoracic duct, all of which must not be touched) in a discectomy intervention which involves removing part of an intervertebral disc.;In this context, it is necessary to calibrate the endoscope using only information contained in the thoracoscopic sequence of images, combined with information from a motion tracking system in order to registrate the peroperative images to a pre-operative 3D model, without requiring any markers on the vertebrae. This calibration procedure consists in determining the intrinsic parameters of the camera (focal length, principal point, zoom) as well as the extrinsic parameters that describe the position of the camera in space throughout the operation.;Minimally invasive surgery reduces the risk of an operation for the patient because the instruments and camera are inserted through small incisions. The camera is used to view the operative field to guide the surgeon's intervention. However, the surgeon loses the notion of depth during such an operation, his vision being limited to a sequence of 2D images acquired by the camera. Therefore, a surgical assistance system could facilitate this type of intervention.;The model of the endoscopic camera differs from the general camera models typically used in the fields of vision and robotics. Thus, existing self-calibration methods for image sequences cannot be used as such.;However, there exist mathematical models describing the formation of endoscopic images. The endoscopic camera possesses additional degrees of freedom that are related to the need in thoracoscopic surgery to have the broadest possible view of the surgical site.;We have thus proposed a self-calibration method for an endoscopic camera whose focal length can vary during the procedure, based solely on the natural content of thoracoscopics image sequences.;To achieve this goal, we started by validating an endoscopic image formation model proposed in the literature. A method for distortion correction and photogrammetric calibration was developed to calibrate the parameters that are not variable during the operation. The distortion is corrected via minimization of two criteria by viewing a regular grid through the endoscope. Photogrammetric calibration is then achieved by basing ourselves on a linear transformation DLT (Direct Linear Transform) using a calibration object consisting of non-coplanar points. Thus, all the intrinsic parameters of the camera as well as its position in space can be initialized before the operation. This method gives good results and has been tested on numerous endoscopic images at several zoom levels.