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

7.2. Methods 113
Additionally, the extracted centerlines are dilated and added to the segmentation result
to assure 6-connected segmentation results.
Parameters: The following set of parameters is used for segmentation of the datasets:
σ = 0.5mm, F max = 200, µ = 5 for the GVF using 500 iterations, t high = 0.5, t low = 0.1,
t s = 5, and t m = 0.5.
7.2.2 Airway Tree Reconstruction Based on Multi-Scale Tube Detection
In this section, we present an automated approach for the reconstruction of airway trees.
It is based no our general approach for segmentation of branched tubular networks as
outlined in Section 1.2 to achieve a high robust against local disturbances which can
result from disease or imaging artifacts, for example. The method consists of three main
processing steps: preprocessing, extraction of tubular objects, and a grouping and linkage
of the identified tubular objects to reconstruct the airway tree. Fig. 7.4 illustrates the
individual processing steps by showing intermediate results.
Preprocessing: To fully automate the approach for airway detection, the input CT
dataset is preprocessed to discard non lung tissue and to restrict the search area for tubular
structures. Therefor, a rough lung mask is generated by using thresholding (< −700HU),
connected component analysis, and morphological closing with a ball structuring element
with a radius of 10 voxels. All voxels outside this lung mask or with a value larger than
−700HU in the original dataset are set to −700HU. The resulting dataset (Fig. 7.4(a))
was used as input for the later steps of the method.
Extraction of tubular structures: For extraction of tubular structures, the tube
detection filter of Pock et al. [113] as described in Section 2.2.2 is utilized in combination
with the height-ridge traversal with hysteresis-thresholding for centerline extraction as
described in Section 2.4. The method of Pock et al. was slightly adapted as for thin
airways the surrounding is not necessarily symmetric and the utilized symmetry criterion
adversely influences the tube detection. Therefore, for these structures (radius ≤ 0.5mm),
the variance-based symmetry criterion is discarded. The tube detection filter response
is shown in Fig. 7.4(b). To discard short spurious responses of the tube detection filter
(noise), all centerlines with an accumulated tube detection filter response below t conf are
discarded. Figs. 7.4(c) and (d) depict the resulting descriptions of the identified tubular
structures. Fig. 7.4(c) shows only the centerline information, while in Fig. 7.4(d) also the