Radiographs of the hand are useful in diagnosing and staging diseases such as
rheumatoid arthritis (RA) and other musculoskeletal diseases. Radiographs are
projections of the 3D anatomy, with the useful information such as pose and pathology becoming lost in the process. Recovering the 3D pose from a single radiograph allows detailed anatomical analysis in the 3D space and possible reduction of radiation exposure for the patients. The research area is around the registration of 3D mesh model to 2D X-ray images for the medical application of automatic disease diagnosis and treatment planning.
The gap in knowledge was that many researches concentrate on registrations of 3D CT volume to radiographs whereas effect methods that work on 3D models are limited.

The aim of this thesis is to develop a 3D pose recovery framework for radiographs using a novel hybrid 3D/2D registration method. Our pose recovery pipeline consists of aligning a simulated digitally reconstructed radiograph of a 3D mesh model to a real radiograph. Our method heavily relies on fast X-ray simulations and optimisation algorithms.
The method used to investigate the feasibility and performance of our approach was through conducting experiments of registrations using synthetic and real clinical data. We evaluated our approach using publicly available datasets including 15 radiographs of hands from the MURA dataset and a radiograph of the hip from the VHP, the Visible Man dataset. Results demonstrated that our approach works well in both registrations of radiographs with two different anatomical structure.

Our key findings were the registration of 3D mesh model to a single radiograph can be performed accurately and fast using our registration framework. Once suitable optimisation algorithm and metrics are identified, our approach tends to be reliable and registration results are similar when repeated for several times. Also, our registration method can be easily modified to solve a specific registration problem where registration of a 3D polygon model to a single 2D image needs to be performed.
The significance of these findings was that it clears the path for further development of a fully automatic process of clinical diagnoses and treatment planning of hand diseases. Further inquiry is required to evaluate the applicability and the performance of our registration approach to other articulated musculoskeletal anatomy, which would lead to the development of clinical applications for diagnosing and treatment planning of any disease targeting a particular anatomical region.