This poster is published under an
open license. Please read the
disclaimer for further details.
Keywords:
Computer applications, Lung, CT, CAD, Computer Applications-Detection, diagnosis, Hypertension
Authors:
N. Kim, S. M. Lee, J. B. Seo; Seoul/KR
DOI:
10.1594/ecr2015/C-0696
Methods and materials
Non-contrast volumetric chest CT scans with sub-millimeter thickness of 10 normal subjects were used.
We extracted pulmonary vessels using a thresholding method (-700 HU) from CT images.
The hilar region was removed from the initial vessel to divide the vessel tree into separated sub-trees.
The skeletons of the tree having mixed artery and vein were extracted using a 3D thinning method,
and minimum spanning tree algorithm were used to differentiate the artery and vein tree separately.
In addition,
the radius estimation on each skeleton voxel was performed using the nearest boundary point,
and the start and end points of each branch was evaluated.
These trees were manually identified as artery or vein by an expert thoracic radiologist[1].
To model these sub-trees,
the root of each sub-tree by using the distance from the hilar region and radius of the each node was automatically determined.
A Dijkstra's algorithm was used to reconstruct each sub-tree into a spanning tree model.
Based on this model,
27 quantitative features including node number,
mean branch asymmetry,
and mean and SD of diameter,
branch order,
branch length,
and tapering ratio,
etc were automatically extracted by using TREES toolbox.
Twenty repetitions with five-fold cross-validation were performed to evaluate overall accuracy of this classifier.