Biomedical Engineering Reference
In-Depth Information
selection and demonstrated that the accuracy of atlas-based segmentation can
be improved substantially by moving beyond the use of a single, individual atlas.
Recently published works on atlas creation and atlas-based segmentation
make increasing use of standard atlases that incorporate properties of a pop-
ulation of subjects [33, 35, 45]. Our results confirm that this is likely beneficial
for improved segmentation accuracy and robustness. However, our results also
suggest that the benefits of applying a multiclassifier strategy are well worth the
extra effort.
On a more application-specific note regarding the accuracy of atlas-based
segmentation of bee brains, we observe that the mean SI value of segmentations
produced using the MUL method in this chapter is 0.86, which, given the small
size of most of the structures in the bee brains considered, is comparable to the
values reported by Dawant
et al.
and supports the visual assessment observation
(Fig. 11.11) that the automatic segmentations described here differ from manual
segmentations on average by slightly more than half of the voxels on the struc-
ture surfaces (Fig. 11.14). In fact, Zijdenbos
et al.
[87] state that “SI
>
0
.
7 indi-
cates excellent agreement” between two segmentations. This criterion (SI
>
0
.
7)
is satisfied by virtually all (97%) contours generated by our segmentations using
the MUL method (Fig. 11.19). Furthermore, since the image quality of confocal
microscopy images is inferior to clinical MR and CT images in many ways, we
believe that our registration-based segmentation method represents a satisfac-
tory intermediate solution to a segmentation problem that is appreciably harder
than that of segmenting commonly used images of the human brain.
Acknowledgments
Torsten Rohlfing was supported by the National Science Foundation under
Grant No. EIA-0104114. While performing the research described in this chap-
ter, Robert Brandt was supported by BMBF Grant 0310961. Daniel B. Russakoff
was supported by the Interdisciplinary Initiatives Program, which is part of the
Bio-X Program at Stanford University, under the grant “Image-Guided Radio-
surgery for the Spine and Lungs.” All computations were performed on an SGI
Origin 3800 supercomputer in the Stanford University Bio-X core facility for
Biomedical Computation. The authors would like to thank Andreas Steege and
Charlotte Kaps for tracing the microscopy images.
