Biomedical Engineering Reference
In-Depth Information
Table 3.
Segmentation Accuracy for Different Numbers of Modes
DSC
D
surface
(mm)
D
centroid
(mm)
3 modes
0.8015
3.55
3.32
10 modes
0.8172
3.38
2.86
25 modes
0.8173
3.46
2.95
While there exist many ways to parameterize this subspace, the representation
in terms of principal components (eigenshapes of the embedding function) has the
additional advantage that the principal components associated with the largest
eigenvalues by definition capture the largest variation of the embedding function.
Hence, one could further reduce the dimensionality of the problem, by usingmerely
the first few eigenmodes.
To quantify the loss in segmentation accuracy when using fewer eigenmodes
in the optimization, we show in Table 3 the values of the Dice coefficient, the
surface distance, and the centroid distance obtained when using 3, 10, and 25
eigenmodes. The reported quantities are averages computed for each of the 25 test
images. As expected, the higher-order eigenmodes contain very little additional
shape information, so that the accuracy increases only by a little amount when
going from 10 to 25 eigenmodes, while the computation time scales linearly with
the number of eigenmodes considered.
6.
CONCLUSION
We proposed herein an efficient and accurate statistical shape prior for level set
segmentation that is based on nonparametric density estimation in the linear sub-
space spanned by the level set surfaces of a set of training shapes. In addition, our
segmentation scheme integrates nonparametric estimates of intensity distributions
and efficient optimization of pose and translation parameters.
We reported quantitative evaluation of segmentation accuracy and speed for
cardiac ultrasound images and for 3D CT images of the prostate. In particular,
we quantitatively validated that the proposed segmentation scheme is robust to
the initialization and robust to noise. Furthermore, we demonstrated that one
can increase efficiency by reducing the number of eigenmodes considered in the
optimization while losing a little accuracy of the average segmentation results.
These results indicate that the proposed nonparametric shape prior outperforms
previously proposed shape priors for level set segmentation.
7.
ACKNOWLEDGMENTS
We thank Christophe Chefd'hotel for fruitful discussions. We thank Marie-
Pierre Jolly for providing us with image and training data for the ultrasound
sequences.
