 |
The development of image-guided surgery techniques provide the
surgeon with intuitive and highly informative intra-operative 3D
visualisations of the internal structures of interest. However, in
order to obtain these 3D representations it is typically required to
obtain 3D image data of the patient (e.g. CT or MRI). In certain
procedures, especially in orthopaedics, there is no routine
acquisition of 3D data, as planning is usually performed in 2D X-ray
images. CT image acquisition induces radiation, and both CT and,
especially, MRI scanning increase significantly the economic
cost. Thus, it is not reasonable to ask the surgeons to take 3D data
solely on the basis that it is needed for the image-guidance.
To address this issue, we have developed methods to estimate the 3D
shape of anatomical structures (bones, in this case) from incomplete
and sparse data about the location of certain points in the
surface. The methods are based on the construction of a model of
normal anatomy, built from a training set of example data. This model
consists of the average bone shape and the principal modes of shape
variation in the training set, built using Principal Components
Analysis. Methods have been devised to fit our model to 3D point data
acquired intra-operatively, either by a tracked pointer or using 2D
ultrasound. Robustness to outliers and small sets of input data is
guaranteed by the incorporation of a novel Mahalanobis shape distance
regularisation term and the use of M-estimators. The system has been
successfully validated by simulation and on cadaver bones.
|
