Biomedical Engineering Reference
In-Depth Information
Figure 7. Prostate segmentation for two patients with the same shape model. Each row
shows axial slices of the same segmentation for one patient. The manual segmentation is
in black and the automatic one white.
false positives increases to 42%. Merely constraining the boundary evolution to
the linear subspace spanned by the training shapes is insufficient to provide for
accurate segmentation results.
5.2. Prostate Segmentation from 3D CT Images
5.2.1. A single statistical shape model for different patients?
Segmentation of the prostate from CT images is an important and challenging
problem in radiotherapy. It may help to avoid the exposure to radiation of vital
organs that are not infected by the cancer. In this image modality, the prostate
appears with an intensity level very close to the one of adjacent organs like the
bladder. The key assumption of our work is that the shape of the prostate in a
given segmentation task is statistically similar to prostate shapes observed in a
training set. Most related works on prostate segmentation are indeed model-based
[7, 10, 11]. In contrast to existing works, we will show that a single (sufficiently so-
phisticated) statistical shape model can be applied to the segmentation of different
patients.
To this end, we built a nonparametric 3D shape model of the prostate using 12
manually extracted prostates (with seminal vesicles) collected from two different
patients.
We employed a leave-one-out strategy by removing the image of interest from
the training phase. Figure 7 shows 2D cuts of a few results obtained using this
strategy. With a one-click initialization inside the organ, the algorithm led to a
steady-state solution in less than 10 seconds. We obtained 86% successfully clas-
sified organ voxels and 11% misclassified organ voxels. This compares favorably
to the intra-patient results reported in [11]. One should note that these quantitative
evaluations underestimate the quality of our results since the “ground-truth” seg-
Search WWH ::




Custom Search