Abstract
This thesis proposes and evaluates new algorithms for segmenting various lung structures in computed
tomography (CT) images, namely the lungs, airway trees and vessel trees. The main objective of
these algorithms is to facilitate a better platform for studying Chronic Obstructive Pulmonary Disease
(COPD) using CT scans from the Danish Lung Cancer Screening Trial (DLCST) study.
We propose a fully automated lung segmentation algorithm that is based on region growing and the
assumption that the lungs are of lower intensities than surrounding structures in CT. Furthermore, we
also propose a post processing step that detects and removes esophagus regions, wrongly added by the
region growing process, to improve the reliability of the lung segmentation algorithm. The proposed
algorithm has been successfully applied to more than 6000 low dose CT scans from the DLCST study.
Among the CT scans applied, 200 randomly selected CT scans were manually evaluated by medical
experts, and only negligible or minor errors were found in nine scans. The proposed algorithm has
been used to study how changes in smoking behavior affect CT based emphysema quantification. The
algorithms for segmenting the airway trees and vessel trees proposed in this thesis are also built on
top of this lung segmentation algorithm.
We propose a voxel classification based airway appearance model for the segmentation of airway
trees, which is trained using easy to obtain manual airway tree segmentations that can be incomplete.
Two approaches for extracting the airway tree using the voxel classification appearance model are
proposed: a vessel guided approach and a locally optimal paths approach.
The vessel guided approach exploits the fact that all airways are accompanied by arteries of similar
orientation. This is accomplished using a vessel orientation similarity measure, which is combined with
the proposed appearance model in a region growing framework. Experiments on CT scans from the
DLCST study have shown that the proposed approach performs better than simply applying region
growing on the response of the appearance model or on the intensity alone. The proposed approach has
also been applied to the diverse set of CT scans from the Extraction of Airways from CT (EXACT’09)
dataset, where the proposed approach has the advantage of having fairly high sensitivity with very
few false positives, when compared to other state of the art airway tree extraction algorithms.
The locally optimal paths approach extracts airway trees by continually extending locally defined
optimal paths, generated using a cost function that incorporates the airway appearance model, and
shape and orientation measures that are derived using a multiscale Hessian eigen analysis. The decisions
in this approach are made on a path basis, which makes it easy to ignore small number of unlikely
airway points on a path. Therefore, the proposed approach is capable of overcoming local occlusions,
which would otherwise stopped region growing based algorithms that make decisions on a voxel basis.
Experiment results have shown that more complete airway trees are extracted with the locally optimal
paths approach as compared to the vessel guided approach, though at a price of a slight increase in
false positive rate. This approach is also used in combination with a multiscale vessel enhancement
filter for the extraction of vessel trees in CT. It was shown that the locally optimal path approach is
capable of extracting a better connected vessel tree and extract more of the small peripheral vessels
in comparison to applying a threshold on the output of the vessel enhancement filter.
Finally, we also constructed a reference standard for evaluating airway tree extraction algorithms
in the EXACT’09 study, which is the first study to perform standardized quantitative evaluation of
different airway tree extraction algorithms based on a standard dataset. Segmented airway trees from
the algorithms that participate in the study were used to construct the reference standard needed,
circumventing the need for labour intensive manual segmentations. Each segmented trees is subdivided
into its individual branch segments, where the branch segments are subjected to visual inspection
by human observers. Branch segments that are determined to be correctly segmented during the
inspection process are then combined to form the reference standard
tomography (CT) images, namely the lungs, airway trees and vessel trees. The main objective of
these algorithms is to facilitate a better platform for studying Chronic Obstructive Pulmonary Disease
(COPD) using CT scans from the Danish Lung Cancer Screening Trial (DLCST) study.
We propose a fully automated lung segmentation algorithm that is based on region growing and the
assumption that the lungs are of lower intensities than surrounding structures in CT. Furthermore, we
also propose a post processing step that detects and removes esophagus regions, wrongly added by the
region growing process, to improve the reliability of the lung segmentation algorithm. The proposed
algorithm has been successfully applied to more than 6000 low dose CT scans from the DLCST study.
Among the CT scans applied, 200 randomly selected CT scans were manually evaluated by medical
experts, and only negligible or minor errors were found in nine scans. The proposed algorithm has
been used to study how changes in smoking behavior affect CT based emphysema quantification. The
algorithms for segmenting the airway trees and vessel trees proposed in this thesis are also built on
top of this lung segmentation algorithm.
We propose a voxel classification based airway appearance model for the segmentation of airway
trees, which is trained using easy to obtain manual airway tree segmentations that can be incomplete.
Two approaches for extracting the airway tree using the voxel classification appearance model are
proposed: a vessel guided approach and a locally optimal paths approach.
The vessel guided approach exploits the fact that all airways are accompanied by arteries of similar
orientation. This is accomplished using a vessel orientation similarity measure, which is combined with
the proposed appearance model in a region growing framework. Experiments on CT scans from the
DLCST study have shown that the proposed approach performs better than simply applying region
growing on the response of the appearance model or on the intensity alone. The proposed approach has
also been applied to the diverse set of CT scans from the Extraction of Airways from CT (EXACT’09)
dataset, where the proposed approach has the advantage of having fairly high sensitivity with very
few false positives, when compared to other state of the art airway tree extraction algorithms.
The locally optimal paths approach extracts airway trees by continually extending locally defined
optimal paths, generated using a cost function that incorporates the airway appearance model, and
shape and orientation measures that are derived using a multiscale Hessian eigen analysis. The decisions
in this approach are made on a path basis, which makes it easy to ignore small number of unlikely
airway points on a path. Therefore, the proposed approach is capable of overcoming local occlusions,
which would otherwise stopped region growing based algorithms that make decisions on a voxel basis.
Experiment results have shown that more complete airway trees are extracted with the locally optimal
paths approach as compared to the vessel guided approach, though at a price of a slight increase in
false positive rate. This approach is also used in combination with a multiscale vessel enhancement
filter for the extraction of vessel trees in CT. It was shown that the locally optimal path approach is
capable of extracting a better connected vessel tree and extract more of the small peripheral vessels
in comparison to applying a threshold on the output of the vessel enhancement filter.
Finally, we also constructed a reference standard for evaluating airway tree extraction algorithms
in the EXACT’09 study, which is the first study to perform standardized quantitative evaluation of
different airway tree extraction algorithms based on a standard dataset. Segmented airway trees from
the algorithms that participate in the study were used to construct the reference standard needed,
circumventing the need for labour intensive manual segmentations. Each segmented trees is subdivided
into its individual branch segments, where the branch segments are subjected to visual inspection
by human observers. Branch segments that are determined to be correctly segmented during the
inspection process are then combined to form the reference standard
| Original language | English |
|---|
| Place of Publication | København |
|---|---|
| Publisher | Faculty of Science, University of Copenhagen |
| Number of pages | 104 |
| Publication status | Published - 2010 |