TY - CONF

T1 - An automatic system for segmentation, matching, anatomical labeling and measurement of airways from CT images

AU - Petersen, Jens

AU - Feragen, Aasa

AU - Owen, Megan

AU - Lo, Pechin

AU - Wille, Mathilde Marie Winkler

AU - Thomsen, Laura Hohwü

AU - Dirksen, Asger

AU - de Bruijne, Marleen

N1 - Poster No.: C-2474

PY - 2013

Y1 - 2013

N2 - Purpose: Assessing airway dimensions and attenuation from CT images is useful in the study of diseases affecting the airways such as Chronic Obstructive Pulmonary Disease (COPD). Measurements can be compared between patients and over time if specific airway segments can be identified. However, manually finding these segments and performing such measurements is very time consuming. The purpose of the developed and validated system is to enable such measurements using automatic segmentations of the airway interior and exterior wall surfaces in three dimensions, anatomical branch labeling of all segmental branches, and longitudinal matching of airway branches in repeated scans of the same subject.Methods and Materials: The segmentation process begins from an automatically detected seed point in the trachea. The airway centerline tree is then constructed by iteratively adding locally optimal paths that most resemble the airway centerlines based on a statistical model derived from a training set. A full segmentation of the wall surfaces is then extracted around the centerline, using a graph based approach, which simultaneously detects both surfaces using image gradients.Deformable image registration is used to match specific airway segments in multiple images of the same subject.The anatomical names of all segmental branches are assigned based on distances to a training set of expert labeled trees. Distances are measured in a geometric tree-space, incorporating both topology and centerline shape differences.Results:The segmentation method has been used on 9711 low dose CT images from the Danish Lung Cancer Screening Trial (DLCST). Manual inspection of thumbnailimages revealed gross errors in a total of 44 images. 29 were missing branchesat the lobar level and only 15 had obvious false positives. A thorough inspection of 10 randomly selected images, revealed the method extracted 174 branches on average and only 3.79% of the found centerline (excluding trachea and main bronchi) to be falsely detected. The extracted wall surfaces were compared to manual annotations in 319 two-dimensional images extracted perpendicularly to and in random positions of the centerline in 7 subjects. Results show an average Dice's coefficient of 89%. The COPD gene phantom was scanned with the DLCST protocol and all interior and exterior diameters were estimated within 0.3 mm of their actual values.Limiting measurements to segments matched in multiple images of the same subject using image registration was observed to increase their reproducibility.The anatomical branch labeling tool was validated on a subset of 20 subjects, 5 of each category: asymptomatic, mild, moderate and severe COPD. The average inter-expert agreement of two trained observers in placing labels L1-L10 and R1-10 was found to be 71%, whereas the system reached 72.7%. Reproducibility of the experts in repeat scans of the same subject, assessed using image registration, was 72.6%, the system reached 76%. Accuracy was not found to be significantly correlated with disease category.Conclusion:The presented system is able to segment the airway wall surfaces in CT images, identify segmental bronchi, and match segments in multiple scans of the same subject. This allows accurate, reproducible and completely automatic analysis of airways in clinical studies of COPD.

AB - Purpose: Assessing airway dimensions and attenuation from CT images is useful in the study of diseases affecting the airways such as Chronic Obstructive Pulmonary Disease (COPD). Measurements can be compared between patients and over time if specific airway segments can be identified. However, manually finding these segments and performing such measurements is very time consuming. The purpose of the developed and validated system is to enable such measurements using automatic segmentations of the airway interior and exterior wall surfaces in three dimensions, anatomical branch labeling of all segmental branches, and longitudinal matching of airway branches in repeated scans of the same subject.Methods and Materials: The segmentation process begins from an automatically detected seed point in the trachea. The airway centerline tree is then constructed by iteratively adding locally optimal paths that most resemble the airway centerlines based on a statistical model derived from a training set. A full segmentation of the wall surfaces is then extracted around the centerline, using a graph based approach, which simultaneously detects both surfaces using image gradients.Deformable image registration is used to match specific airway segments in multiple images of the same subject.The anatomical names of all segmental branches are assigned based on distances to a training set of expert labeled trees. Distances are measured in a geometric tree-space, incorporating both topology and centerline shape differences.Results:The segmentation method has been used on 9711 low dose CT images from the Danish Lung Cancer Screening Trial (DLCST). Manual inspection of thumbnailimages revealed gross errors in a total of 44 images. 29 were missing branchesat the lobar level and only 15 had obvious false positives. A thorough inspection of 10 randomly selected images, revealed the method extracted 174 branches on average and only 3.79% of the found centerline (excluding trachea and main bronchi) to be falsely detected. The extracted wall surfaces were compared to manual annotations in 319 two-dimensional images extracted perpendicularly to and in random positions of the centerline in 7 subjects. Results show an average Dice's coefficient of 89%. The COPD gene phantom was scanned with the DLCST protocol and all interior and exterior diameters were estimated within 0.3 mm of their actual values.Limiting measurements to segments matched in multiple images of the same subject using image registration was observed to increase their reproducibility.The anatomical branch labeling tool was validated on a subset of 20 subjects, 5 of each category: asymptomatic, mild, moderate and severe COPD. The average inter-expert agreement of two trained observers in placing labels L1-L10 and R1-10 was found to be 71%, whereas the system reached 72.7%. Reproducibility of the experts in repeat scans of the same subject, assessed using image registration, was 72.6%, the system reached 76%. Accuracy was not found to be significantly correlated with disease category.Conclusion:The presented system is able to segment the airway wall surfaces in CT images, identify segmental bronchi, and match segments in multiple scans of the same subject. This allows accurate, reproducible and completely automatic analysis of airways in clinical studies of COPD.

U2 - 10.1594/ecr2013/C-2474

DO - 10.1594/ecr2013/C-2474

M3 - Poster

Y2 - 7 March 2013 through 11 March 2013

ER -