Deformable Models in Medical Image Segmentation - 3D Multiscale Physiological Human - page 77

Deformable models have a broad range of possible applications. They have been
used in computer graphics [ 3 ], to calculate the deformation of clothes [ 4 ] and in
image composition [ 5 ]. Other uses include optical flow analysis for facial animation
[ 6 ] and the determination of vehicle types [ 7 ].
Due to their power and robustness, deformable models are often used in medical
applications. For example, Deserno et al. [ 8 ] presented an application for segment-
ing the bony orbit while Heimann et al. [ 9 ] proposed a grand challenge for knee
segmentation. Schmid [ 10 ] presented the segmentation of the hip bone, Snel et al. [ 11 ]
used a deformable model for wrist segmentation and Rafai et al. [ 12 ] proposed a
method for skull segmentation. Another medical use is surgery simulation which
has been reviewed by Meyer et al. [ 13 ]. Medical simulation approaches that use
deformable models to describe the mechanical behavior need proper evaluation.
Marchal et al. [ 14 ] proposed a new and open framework to combine several metrics
and models to compare different algorithms.
Several surveys on deformable models have been published, e.g. by McInerney
et al. [ 15 ] in 1996, Montagnat et al. [ 16 ] in 2001 and Hegadi et al. [ 17 ] in 2010. Moore
et al. [ 18 ] provided a general survey on deformable models while Jain et al. [ 19 ]
have reviewed deformable template models, which are used for segmentation, image
retrieval and video tracking.
Deformable models can be divided into discrete and continuous representations
[ 16 ]. The discrete models can be split into particle systems and meshes. Contin-
uous representations can have either an explicit (snakes) or implicit (level sets)
description of their surface. These different classes of deformable models will be sur-
veyed in this overview. First we will introduce active contour models, also known as
snakes (Sect. 4.2 ). Next, several approaches to the level sets method will be discussed
(Sect. 4.3 ), followed by a description of discrete deformable models (Sect. 4.4 ).
We take a look at knowledge-based deformable models (Sect. 4.5 ) that are a spe-
cialization and at some alternative approaches (Sect. 4.6 ). An overview of external
forces (Sect. 4.7 ) and approaches for segmentation initialization will also be given
(Sect. 4.8 ). In the conclusion, the most important aspects are summarized and future
prospects are briefly discussed.
4.2 Snakes
Active contour models, usually called snakes, are a class of deformable models
with a continuous representation based on an explicit surface description. They have
been originally proposed by Kass et al. [ 20 ] in 1988 and have been applied as a
method for edge detection [ 21 ], motion tracking [ 22 ], stereo matching [ 23 ] and
interactive image interpretation [ 20 ]. Other applications are shape modeling [ 24 , 25 ]
and segmentation [ 21 , 26 ].
The main concept is to use an energy-minimizing parametric curve or a spline,
reducing the problem to aminimization problem. Snakes need to be initialized closely
to the final shape to ensure that they are attracted to the correct features.