Statistical Shape Models, Segmentation, Landmark detection, CBCT , Inferior alveolar nerve injury
JudChristoph, LüthiMarcel, AlbrechtThomas, Schönborn Sandro, Vetter Thomas (2014), Variational Image Registration Using Inhomogeneous Regularization, in Journal of Mathematical Imaging and Vision
, 48, 1-15.
Lüthi Marcel, Jud Christoph, Vetter Thomas (2011), Using Landmarks as a Deformation Prior for Hybrid Image Registration, in DAGM'11: 33nd Annual Symposium of the German Association for Pattern Recognition
, FrankfurtSpringer, Berlin / Heidelberg.
Lüthi Marcel, Jud Christoph, VetterThomas (2013), A unified approach to shape model fitting and non-rigid registration, in Prceedings of the 4th International Workshop on Machine Learning in Medical Imaging, LNCS 8184
, NagoyaSpringer International Publishing, Cham.
Jud Christoph and VetterThomas (2014), Using object probabilities in deformable model fitting, in International Conference on Pattern Recognition (ICPR)
, IEEE, USA.
This project is a collaborative approach of the Computer Science Department and the Department for Oral Surgery at the University Basel.
Surgical extraction of wisdom teeth (third Molar) is the most commonly performed procedure in oral surgery. Depending on the anatomical position and the distance of the roots to the mandibular canal, this procedure implies the risk of injury of the inferior alveolar nerve. In the worst case this leads to a permanent loss of sensation in the lower lip and the chin. With the introduction of Cone Beam Computed Tomography (CBCT), a new imaging technology is available, which allows to obtain a detailed representation the patients anatomy with minimal radiation dose. The resulting three dimensional images open new possibilities in the risk assessment and planning of these surgeries.
The goal of this project is to develop a method for automatic segmentation of the wisdom tooth and the mandibular canal from the CBCT images. In parallel, the indicators for an injury of the inferior alveolar nerve given in these images are studied by experienced oral surgeons. The pre-operative risk assessment of the surgeons will be compared with the actual outcome of the surgery. Based on the experience of this study a software for automatic risk assessment from the images will be developed and evaluated on the collected data.
For being able to perform an automatic risk analysis, we plan to perform segmentation of the mandibular canal and the wisdom tooth. Besides allowing for a risk assessment, having a segmentation of these tissues is of independent interest for the clinical application. It allows to generate a three-dimensional model of the detailed patient's anatomy which can support the surgery planning and serve as a visual aid for patient education.
Automatic segmentation of the wisdom tooth from CBCT is made difficult by the low contrast, which makes the distinction of different tissue types difficult, and the fact that individual teeth may be touching. We approach both problems by fitting a statistical shape model, which restricts the segmentation result to the normal anatomy of the tooth. A further complication which arises is that the the spatial orientation of the tooth can greatly vary. To address this issue we develop an algorithm for detecting feature points in the images, such as the tips of the roots and the center of the crown, which allow us to determining its spatial alignment. The same detection algorithm will be used for detection of the mandibular canal. The canal itself is not clearly visible in the images, but at some places the cortical rim around this canal can be detected. Given the information of the exact position at some points, we will use shortest path algorithms, with edge-weights depending on the image intensities, to determine the exact course of this canal.
We see the following main challenges in our approach: The variability of the wisdom tooth is extremely large and even topological changes can occur. Therefore, the standard approach for statistical shape models cannot be used for accurately representing this tooth. We will need to be able to flexibly switch between different models, based on the topology of the target shape. Another challenge is that the spatial alignment of the wisdom tooth can greatly vary. This implies that the algorithms we use for detecting the feature points have to be rotational invariant. Furthermore, to be able to efficiently find feature points in three-dimensional images, a sophisticated classification strategy has to be employed.