We have presented an adaptation and extension to the geometric deformable model. By introducing a novel means of identifying the object boundary and introducing a phase of relaxation to the model deformation we can avoid the problems of surface self intersection, local minima and leakage.
Probabilistic labelling provides the information needed for the deformable surface to identify the object boundary. This information is derived from the histogram of intensities in the image volume by modelling the intensity distributions for the feature and non-feature as Gaussian distributions. These Gaussians can then be used to produce a likelihood that a given voxel with a particular intensity belongs to the feature being segmented.
Relaxing the model during the deformation stage releases nodes trapped in local minima by imposing strict rules on smoothness. We also identify nodes that appear to be leaking out of the feature boundary and force them to contract back to the mesh. The deformation process of ballooning and relaxation may take longer to complete than conventional models but does not require any user-interaction to obtain an accurate model.
The purpose of this work is to create an automatic means of assessing and monitoring the progress of Rheumatoid arthritis in knee joints. The automatic segmentation of the femur is the first stage in this assessment, and our method produces a model which approximates the surface of the femur. Future work will involve investigating the feedback of information obtained from the segmentation into the probabilistic labelling. The next stage of the assessment will be the automatic segmentation of other anatomy in the knee such as the patella.
N Hill