Bastian Bier
C-arm Cone-Beam Computed Tomography Reconstruction for Knee Imaging under Weight-Bearing Conditions
Medical imaging is essential for the assessment of osteoarthritis and the overall knee health. For that purpose, radiographs of the knees of standing patients are acquired commonly. These suffer, however, under projective transformation and thus do not allow conclusions to be drawn about the complex 3-D joint anatomy. Conversely, compared to many 3-D capable imaging modalities, imaging under load is easily feasible using X-rays. This is beneficial, since it has been shown that this reflects the knee joint under more realistic conditions. Recently, a 3-D imaging protocol has been proposed that enables cone-beam computed tomography (CBCT) reconstruction of the knees acquired under weight-bearing conditions. To this end, the C-arm rotates on a horizontal trajectory around the standing patient. Involuntary patient motion and scattered radiation deteriorate the reconstructions’ image quality substantially. In this thesis, novel concepts and methods are proposed to further develop this imaging protocol in order to improve the reconstructions. In a first approach, a primary modulator-based scatter correction method has been transferred on a clinical C-arm CBCT system. The method is a suitable candidate to be applied to projection images of the knees, since it is capable of estimating heterogeneous scatter distributions. To this end, extensions to an existing method have been developed to compensate for the system wobble and the automatic exposure control of the C-arm systems. In multiple experiments, it is demonstrated that the primary modulator method works on clinical C-arm scanners and also for imaging under load. A current state-of-the-art motion estimation method is based on metallic fiducial markers that are placed on the patient’s skin. The marker placement is, however, a tedious and time-consuming process. To this end, two marker-free alternatives are proposed. In a first attempt, a range camera is utilized to track the patient surface simultaneous to a CBCT image acquisition. Using point cloud registration of the acquired depth frames, transformations can be computed that correspond to the patient motion, which directly can be integrated into the image reconstruction. In a simulation study, comparable results to the marker-based method could be achieved. Yet, initial real data experiments on a clinical scanner did not achieve satisfying image quality, even though part of the motion could be estimated. Therefore, the promising results make this method to a pre-cursor to future research. Although this method is marker-free, a prepared environment is required. Hence, another purely image-based motion estimated approach has been investigated. The idea is to replace the positions of the fiducial markers with the ones of anatomical landmarks present in the projections. Anatomical landmark detection in X-ray images from different directions is difficult and, to the best of our knowledge, has not been investigated yet. For this purpose, a novel deep learning-based approach has been developed. In a first evaluation, the method was tested on X-ray images of the pelvis. Here, it could be demonstrated that the detection accuracy sufficed to initialize a 2-D/3-D registration. Subsequently, the approach is transferred to knee projection images, where the good detection results served as input for the motion estimation. Despite limited results on real data acquisition, the achieved improvements of the image quality are an indicator for a successful future application for motion estimation.