Biomedical Engineering Reference
In-Depth Information
1.
INTRODUCTION
As medical imaging plays an increasingly prominent role in the diagnosis
and treatment of diseases, thousands of CT, MR, PET, and other images are ac-
quired daily [1]. Their efficient and fast processing is challenging. Automatic or
semiautomatic segmentation of different image components is useful for analyz-
ing anatomical structures, the spatial distribution of their function or activity, and
pathological regions. Segmentation is the basis for visualization and diagnosis.
Many segmentation methods have been proposed for medical image segmen-
tation. Model-based methods utilize training in advance of obtaining sufficient
prior knowledge, in order to get a fine parametric representation. However, these
methods can only deal with one structure at a time. Although Snake-driven ap-
proaches [2, 3] have been successful in segmenting medical images that cannot be
segmented well with traditional techniques, these approaches suffer from the strict
requirement of initial contour and high computational load. In our work, we use
level set methods as the tool, due to their strength in dealing with local deforma-
tions, multicomponent structures, and adaptability of topology. Such capabilities
make these methods suitable for segmentation of medical images, which typically
contain serial images of internal tissues with topological changes.
In the level set method, the construction of the speed function is vital with
respect to the final result. The speed function is designed to control movement
of the curve; and in different application problems, the key is to determine the
appropriate stopping criteria for the evolution [4]. In our work, two improvements
for enhancing the speed function of level set methods are presented, to tackle the
segmentation of the left ventricle from taggedMR images. Both improvements are
related to the construction of the speed function, which controls the propagating
interfaces for the purpose of segmentation. Our method is especially efficient in
segmenting the tagged MRI images with blurry boundaries and strong tag lines.
Two ways to enhance the speed function for segmentation of images with blurry
boundaries and strong tag lines are presented here. We first introduce the
relaxation
factor
. It provides a way to stop the evolving curve at the ventricle boundary
even when the boundary is blurry. In other words, the relaxation factor relaxes
the bounding condition of the difference between the maximum and minimum
gradient values of the image. This approach is described in detail in Section 4.1.
This initial idea solves the leak problem at the blurry boundary by relaxing the
bounding condition of the speed function. Second,
image content
-based items are
incorporated in the speed function. The goal here is to take care of more image
properties in order to control the front propagation instead of using the image
gradient alone. We derive a simple and general model, through combining the
image content items, to endow the speed function with more variability and better
performance. By exploiting more image features than the image gradient, the
constructed speed function can force the evolving boundary to stop at the object
boundary, even if strong tag lines are close to the true boundary.
