Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE credits
A central problem in cranio-maxillofacial (CMF) surgery is to restore the normal anatomy of the facial skeleton after defects, e.g., malformations, tumours, and trauma to the face. There is ample evidence that careful pre-operative surgery planning can significantly improve the precision and predictability of CMF surgery as well as reduce the post-operative morbidity. In addition, the time in the operating room can be reduced and thereby also costs. Of particular interest in CMF surgery planning is to measure the shape and volume of the orbit (eye socket), comparing an intact side with an injured side. These properties can be measured in 3D CT images of the skull, but in order to do this, we first need to separate the orbit from the rest of the image—a process called segmentation.
Today, orbit segmentation is usually performed by experts in CMF surgery planning who manually trace the orbit boundaries in a large number of CT image slices. This manual segmentation method is accurate but time-consuming, tedious, and sensitive to operator errors. Fully automatic orbit segmentation, on the other hand, is unreliable and difficult to achieve, mainly because of the high shape variability of the orbit, the thin nature of the orbital walls, the lack of an exact definition of the orbital opening, and the presence of CT imaging artifacts such as noise and the partial volume effect.
The outcome of this master's thesis project is a prototype of a semi-automatic system for segmenting orbits in CT images. The system first extracts the boundaries of the orbital bone structures and then segments the orbit by fitting an interactive deformable simplex mesh to the extracted boundaries. A graphical user interface with volume visualization tools and haptic feedback allows the user to explore the input CT image, define anatomical landmarks, and guide the deformable simplex mesh through the segmentation.
To evaluate the performance of our segmentation system, we let three test users independently segment 14 orbits twice (in a set of seven CT images) with the segmentation tools provided by the system. In order to assess segmentation accuracy, we construct crisp and fuzzy ground truth segmentations from manual orbit segmentations performed by the three test users. The results of this case study indicate that our segmentation system can be used to obtain fast and accurate orbit segmentations, with high intra-operator and inter-operator precision.