Image Processing Reference
In-Depth Information
from examples. In this area, several recent efforts have shown promising
results [146,152,153,172, 185,186,193,217,231] and new techniques are
likely to continue to appear in the near future. These modeling approaches
go from shape examples to a specific shape representation; they can
reduce computational demands and improve robustness. A small number
of efficiently selected model parameters reduces the dimensionality of
the model recovery problem, and naturally constrains its results owing
to model specificity.
Further investigation of suitable image features will be needed to
improve shape recovery. In particular, incorporation of domain knowl-
edge about the type of image modality (and acquisition protocols) can
play an important role in increasing the accuracy of shape recovery
techniques. In this area the use of image registration techniques is
assuming increasing importance in cardiac image analysis [255], given
that it facilitates the fusion of information from multiple modalities
into a single model reference frame [171].
Most of the initial modeling techniques presented in this review were
either purely geometric or inspired by a virtual physical analog (phys-
ics-based approaches). More recently, several papers have introduced
the known biomechanical properties of the heart in the formulation
of models that analyze cardiac images [167,171,190,191,243]. Fur-
ther development of such approaches, and their application to seg-
mentation tasks, can be a natural way of extending the ideas of
physics-based methods and of relating some of the
ad hoc
parameters
with the experimental evidence provided by biomechanics. Combi-
nation of other physical phenomena such as electromechanical cou-
pling into image-based analysis has also been explored by some
authors [234].
2.
Research on interactive model-based segmentation:
Table 9.2
supports
the idea that model-based cardiac segmentation has not reached the
status of being effectively automated because current techniques either
require substantial expert guidance,
ad hoc
parameter fine-tuning, or
nontrivial preprocessing. Although full automation is a desirable end
goal, its difficulty has been acknowledged many times in the literature.
There is a growing consensus that user interaction is, to some extent,
unavoidable, and that it has to be considered as an integral part of the
segmentation procedure. Therefore, development of efficient tools for
3-D interaction will play an important role in the near future. Being
efficient entails the operator keeping control over the segmentation
process to correct it or overrule its results where it has failed, with
minimal and intuitive user interaction, and guiding the algorithm in
abnormal situations (e.g., in dealing with a pathological case). Of
course, the issue of reproducibility in cases of human intervention
needs attention. Where well-defined repetitive tasks are recognized, or
where a local user interaction can be extrapolated to a broader area,
Search WWH ::
Custom Search