Segmentation steps:
Figure 1: (a) Sagittal slice #-24. (b) Coronal slice #16. (c) The cross-section of slices from figures (a) and (b), containing the seed point and six locally-maximum gradient points.
From a seed point selected by the user, the algorithm will search for six points on the boundary of the tumor. The average intensity of these points will provide the initial threshold value.
 |
 |
| Figure 2: Image thresholded at T=120 obtained from the process shown in Figure 1. | Figure 3: Result of removing unnecessary surfaces from Fig. 2 that did not contain the seed point. |
The volume is thresholded at the obtained threshold value T and removing regions that did not contain the seed point.
 |
 |
| Figure 3: Image thresholded at T1=99 with unnecessary regions removed. | Figure 4: Image thresholded at T2=148 with unnecessary regions removed. |
 |
 |
 |
| Figure 5: Different cross-sections from combining images from figures 3 and 4. The white area defines the mask that is used to filter out irrelevant edges. | ||
By displacing the initial boundary obtained from step 2 by a distance d, two threshold values T1 and T2 are obtained. The result of thresholding the volume at T1 and T2 can be combined as shown in Fig. 5 to define a mask and filter out irrelevant edges.
 |
 |
| Figure 6: Zero-crossing edges after removing 90% of the smallest magnitude edges. | Figure 7: Image edges falling inside the mask provided by step 3. |
Zero-crossing edges are computed from the volumetric image. Only edges falling in the top 10% of the gradient histogram are considered. Figure 7 shows only the tumor-relevant edges falling inside the mask provided by double thresholding. A search for tumor boundary edges depicted in Figure 7 will be conducted next.
Figure 8: Edge points resulting from the nearest-neighbor search in the edge image of Figure 7.
For every point on the boundary from Figure 3, an edge will be assigned from Figure 7. By the nearest-neighbor approach, the closest edge point will be selected. Figure 8 shows the resulting edges after the search. In order to provide a closed tumor surface, an elastic surface is fitted to the edge points.
Figure 9: Three-dimensional view of the elastic Gaussian surface (sigma=3.0) fitting to the edge points.
 |
 |
| Figure 10: Coronal cross-section of the segmentation result. | Figure 11: Sagittal cross-section of the segmentation result. |
Figures 10 and 11 show two slices of the segmentation result after fitting a RaG surface to the edge points (Fig. 9).
[Go Back] [CSE Department Home Page] [Intelligent Systems Lab Page]
For more information contact: A. Goshtasby (ardeshir@cs.wright.edu).
Last modified: 8/21/97.