Biomedical Engineering Reference
In-Depth Information
Figure 10.
(a) The same example as in Figure 8d; (b) after applying stick detection using
5
×
5
masks.
5.2. Segmentation Subsystem for Ultrasound Image Analysis
The level set method (see Sethian and colleagues [19,32]) is a numerical
technique for computing and analyzing front propagation. It offers a highly robust
and accurate method for tracking interfaces moving under a regime of complex
motion. The main idea was introduced in Section 2. During the preprocessing
procedure for the eight continuous images extracted from the dataset, we applied
2D anisotropic filtering and a 2D stick-based enhancement method to each image.
Because the 2Dmethod is inadequate for calculating the volume of sideslip during
compression, the 3D level-set method was used for continuous ultrasound images.
Furthermore, changes in the neighboring 2D images were very small and the
neighboring images very similar, so that the time axis
t
can be considered as the
z
axis in 3D coordinates. With the information of the previous and next frames, the
3D level set method can yield amore accurate segmentation than the 2Dmethod. In
Figures 11 and 12, two segmentation results for different tumor cases, both benign
and malignant, are illustrated, respectively. It should be noted that malignant cases
can be segmented as well as the benign ones.
In order to emphasize the importance of image processing, we compare using
the level set method on a non-enhanced image and an enhanced image, as shown
in Figure 13. According to the results depicted in Figures 11 and 12, our expec-
tations in terms of accuracy of segmentation were satisfied for both the benign
and malignant cases. However, the significance of these preprocessing techniques
were still not explicit. Therefore, using the same examples as in Figures 11 and
12, the original ultrasound images were directly processed using the level set ap-
proach. Because speckle and noise affect the quality of the ultrasound images,
the segmentation results in the middle column of Figure 13 do not seem superior.
This shows the importance of the effect of anisotropic diffusion filtering and stick
detection.
