Segmentation of 3D Tubular Tree Structures in Medical Images ...

94 Chapter 5. Liver Vascular Tree Segmentation
ven vessel trees on clinical CT datasets (Section 5.3.2) and compared it to other methods
with similar objectives (Section 5.3.3). The method performs well and produced only a
few minor errors without clinical significance, contrary to the other methods. In particular,
the structural grouping and linkage with the incorporated flow direction information
was an important factor in obtaining the correct structure of multiple overlapping (vessel)
trees (Sections 5.3.2, 5.3.3, and 3.4.1)
Segmentation accuracy: On phantom datasets we quantified the segmentation accuracy
for tubular objects with different radius under varying imaging conditions (Section
5.3.1). Investigated contrasts and scan resolutions showed almost no effect on the
accuracy in case of vessels with larger diameter. For thin tubular objects the statistics
was not very meaningful, because only a few tube elements were detectable (about 90%
undetected) due to low contrast, noise, and low resolution. For all successfully identified
vessels, the absolute radius error stayed within an one voxel range (inter-slice resolution).
Note that the used graph cut segmentation is only able to produce voxel accurate segmentations.
We also scored the segmentation accuracy on liver CT datasets in terms of
clinical usability for surgical planning (Section 5.3.2). The majority of segmented branches
were scored as ”good“ and no branch was scored as ”poor“, even for datasets with ”poor“
data quality. For low quality datasets with only few visible generations, the segmentation
tended to be scored as ”ok“ toward the distal parts of the vessel trees where the contrast
almost vanishes.
Robustness: Our method performed robustly on the clinical datasets of the liver CT
data as shown in our evaluation. For example, it produced correct results in disturbed
regions caused by adjacent tumors (Figs. 5.9 and 5.12). All liver datasets utilized in our
evaluation contained multiple overlapping vessel trees that had to be separated, and 11
out of the 15 datasets had pathological variations where other methods are likely to fail,
as discussed in Section 5.3.3.
5.5 Conclusion
In this chapter, we presented and validated an approach for the segmentation and separation
of the livers blood vessel trees. In contrast to other approaches, our method
performs an identification of tubular objects followed by a a structural analysis to obtain
the structure of the different vessel trees. This structure information is then utilized as