Biomedical Engineering Reference
In-Depth Information
Figure 26. The region-based energy is decreasing as the Rprop method is applied for the
2D model-based segmentation.
5.3. Extend to Higher-Dimensional Problems
From the previous section, solution for the 2D segmentation using a shape
model was introduced. As the imaging technology is becoming more and more
powerful in clinical usage and medical research, 3D volumetric datasets or even
4D image datasets are used more often in diagnosis. Higher-dimensional imaging
has multiple advantages over 2D imaging in many aspects. It can provide accurate
quantification of the volumes and shape for the organ under observation without
the need for geometric assumptions. It may also improve visualization of spatial
relations between structures at different phases that are not readily obtained from
conventional 2D images.
It is not a complicated task to extend the 2D version of the shape-based seg-
mentation using level set methods to a higher-dimensional space. In the following,
a 3D example is illustrated that is from synthesized 3D images. In Figure 27, five
3D images are listed that are used for training. The size of the image is 100
×
100
100 voxels. In Figure 28, six new shapes are generated by using the 3D
eigenvalues after PCA in the training stage.
After 3D training, we applied the shape information to 3D segmentation. In
Figure 29 the reader will find an extremely noised 3D image where a dumbbell-
like shape is on the inside. Initially, the average shape is placed in the center of
the 3D image. Then 3D model-based segmentation is applied. Three key steps
illustrated in Figure 29. The red contours in the left column correspond to the
current segmentation. The left column is the surface rendering of the current
×
Search WWH ::




Custom Search