The 3D reconstruction of scenes using monocular video sequences is a classical problem in computer vision. The literature covers, not only 3D reconstruction, but also related topics as camera calibration and 3D registration. However most of the described methodologies were developed for perspective images of everyday scenes. The goal of this project is to extend this framework to endoscopic video sequences in the context of computer aided orthopedic surgery (CAOS). Our target application is the visualization and navigation during the reconstruction of the Anterior Cruciate Ligament (ACL) using arthroscopy. Since we aim to develop methods to extract 3D information from endoscopic images, the project outcome will be useful for many other minimally invasive medical procedures.
(1) harvest a tendon graft from the patellar or hamstring;
(2) remove the torn ligament;
(3) open drill holes in the tibia plate and femur notch at the attachment site of the original ligament;
(4) pull the graft through the open tunnel and into the knee joint;
(5) held the tendon graft in place with screws.
The ACL is the major stabilizing ligament of the knee. It is located in the center of the knee joint and runs from the femur to the tibia. The tearing of the ACL occurs with a sudden direction change or when a deceleration force crosses the knee. It usually requires surgical treatment where the torn ligament is reconstructed through minimally invasive procedure. The steps for the arthroscopic ACL reconstruction are the following:
Video of the opening of the femural tunnel
Video of the opening of the tibia tunnel
The procedure must be performed by well trained surgeons. The surgeon must be able to navigate in the knee joint, recognize anatomical landmarks, and open the graft tunnel using only the visual feedback provided by the arthroscope. The accurate positioning of the drill holes from the outside is the most critical and difficult step. Small errors in the graft placement can lead to abnormal tensions during motion, cause pain and new injuries, and ultimately force to a correction surgery. Studies show that the rate of clinical success is 85%, and that 50% of the failures are due to graft misplacement. Certain authors claim that significant errors in positioning the tunnel occur in 10 to 40% of the ACL reconstructions. This is a scenario where CAOS can have strong social and economical impacts. Computer systems for enhancement of surgeon's perception and precise navigation in the knee joint can improve the clinical success rate and decrease the practitioner learning curve. Since the tearing of the ACL is a common injury (over 75000 cases per year just in the US), the development of such systems can bring great benefits in improving the life quality of young patients and diminishing health care costs.
A surgical navigator is a visualization system providing real-time positional information about instruments and tools with respect to a target organ (bones). The position and orientation of external rigid objects, like the arthroscope and surgical tools, can be easily computed using a commercial optical tracker. A stereo head tracks a set of LED markers that are rigidly attached to each instrument. Simple triangulation is used to determine the 3D pose of the different tools in a common world reference frame. The main difficulty in developing a CAOS system for ACL reconstruction is the estimation of the position and orientation of the femur/tibia. The patient's leg moves during the surgery, and optical tracking can not be used because the bones are not visible from the outside. We propose to solve the problem by reconstructing the knee joint from the arthroscopic video stream. The partial reconstructions can be registered with pre-operative volumetric images (typically CT scan) in order to estimate the 3D pose of the femur/tibia with respect to the camera. Since the rigid motion of the arthroscope is tracked by the external stereo head, we can compute the position and orientation of the bones in the world coordinate system where the surgical tools are referenced.
Despite of the recent developments in computer vision, working with arthroscopic images remains a challenging problem. We will focus our research efforts in the following open issues:
1. Geometry of Image Formation - The arthroscope is a non-conventional vision sensor and can not be described by the perspective model. We intend to derive a suitable projection model, study the image geometry, and develop calibration methods.
2. 3D Reconstruction - The reconstruction of the knee joint is difficult due to the nature of the scenes, the image distortion and the lightning conditions. The camera calibration and the tracking of its motion will partially constraint the problem. Nevertheless we need to investigate features types that can be reliably tracked, deal with textureless regions, etc.
3. 3D Registration - The recovered local structure must be registered with a pre-operative CT scan. The 3D registration algorithm must be accurate despite of the leg's movements, sparse local 3D information, and real-time requirements.