Abstract
This thesis investigates an accurate method for three dimensional (3D) reconstruction of the human spine from bi-planar radiographs with comparable results to CT scans or MRI. In this work, we generated a publicly available dataset which corresponds to the training data used. We subsequently solved the problem of correspondences using a landmark-free algorithm applied on the vertebrae. Finally, we developed a semi automatic method based on simulated radiographs for the reconstruction of the human lumbar spine in 3D from bi-planar radiographs.
We validated the results in vitro on radiographs of dried vertebrae with models constructed from a laser-scanner, then in vivo on radiographs of living patients with models extracted from CT scans or MRI.
The results show the feasibility of generating personalised models of patients from bi-planar radiographs.
The contributions of this thesis are:
- Evaluation of the methods for creating 3D models of vertebrae and estimation of the errors in comparison with ground truth data. These methods are applicable to other free-form shapes;
- Creation of landmark free ASMs of lumbar vertebrae;
- Definition and evaluation of a process for estimating the shape and position of lumbar spine from uncalibrated bi-planar radiographs.