Novel 3D statistical shape models for segmentation of medical images
Zheen Zhao · 2006
snakes.This is because the SSMs allow themselves to deform according to the shapes of objects as long as their deformed shapes falling into the plausible area extracted from training sets.Moreover, SSMs incorporate segmentation and interpretation as one integrated operation.It avoids the information loss which occurs between the steps of conventional processing.All the advantages of SSMs have motivated the author to further investigate their usage in the area of medical image analysis, with the emphasis on 3D segmentation.However, automated construction of models is still an open topic.The plausible area/allowable region of a SSM is constructed from a Point Distribution Model (PDM) which is developed from a set of consistent points known as landmarks.Landmarks should represent all training shapes consistently, efficiently and accurately.Consistency in this context means that each landmark corresponds to approximately the same contour/mesh location on all samples.Denoting landmarks along contours is tedious and subjective even in the case of the 2D model.Manual construction of 3D PDMs is highly impractical because of the large number of landmarks.It is typically of the order of thousands.Moreover, 3D SSMs often restrict themselves from detecting details during deformation.The reason is that the dimension of the model is of typically two or three orders of magnitude higher than the number of training samples.It is difficult to estimate a high-dimensional probability distribution of features from a relatively small number of samples [18].To solve these problems, the focus of this thesis is to develop automated methods for the construction of PDMs, as well as new SSMs which can accurately segment objects as only small training sets are available.