Biomedical Engineering Reference
In-Depth Information
4.5 Knowledge-Based Deformable Models
Deformable models require initial shape information to achieve better results.
Knowledge-based deformable models are a group of approaches that aim to invoke
more prior knowledge through additional features. This increases the flexibility in
poor images and the robustness against inaccurate initializations. A comprehensive
overview of knowledge-based deformable models has been given by Schmid [
10
]
in 2011. A simple approach would be assigning labels with an associated behavior
to the model. Other models use the statistical or probabilistic variation of selected
features. In this section, possibilities for feature selection are discussed, followed by
the alignment process and construction of the statistical model. We conclude with the
description of the two main approaches: active shape models and active appearance
models.
4.5.1 Feature Selection
Features select the property of the models that will be exploited. They can be summed
up in three categories: shape-, appearance- and transformation-based features.
4.5.1.1 Shape-Based Features
Cootes et al. [
70
,
71
] have proposed shape-based features in 1994. They use prior
shape information to improve results and limit shape variability to the shape variations
from training data. A set of points, called landmarks, represents image features. They
have to be (usually manually) selected from a large number of training data. Another
approach for these point distributionmodels (PDM) can be used to replace landmarks
with parameters of a medial axis [
72
].
4.5.1.2 Appearance-Based Features
To exploit image properties, appearance-based features can be intensity, gradient,
texture, momentum, etc. Intensity profiles (IP, see Sect.
4.7.2.1
) are very important
and commonly used. In 3D, appearance-based features suffer from an increased
complexity (highly increased memory consumption and computational effort).
