Lung and Lung Lobe Segmentation Methods at Fraunhofer MEVIS

In this paper we present one method for segmenting the lungs and three methods to segment pulmonary lobes from thoracic CT images and their application to the LOLA11 challenge data. The lung segmentation procedure is fully automated and uses a sequence of morphological operations to refine an initial threshold-based segmentation of the pulmonary airspaces. Based on its results, lobe segmentation is performed. The three presented lobe segmentation methods differ substantially in grade of automation. The first lobe segmentation method is a fully automatic segmentation algorithm that combines information from lobar fissures, blood vessels and the airway tree by means of a watershed transformation. The second presented method describes an efficient interactive correction mechanism for existing lobe segmentations. The user can iteratively modify a lobar boundary by drawing its correct course onto regions of insufficient segmentation, getting instant feedback of the results of his actions. The third presented algorithm is an interactive method related to the second one, but it allows for segmentation from scratch based on a lung mask only. Evaluation of the methods was performed as part of the LOLA11 challenge on 55 CT scans that can be considered challenging due to a large number of substantially abnormal cases. Automated lung segmentation took 1minute on average and the mean overlap with the reference standard was 97.3%. For the fully automated, interactively corrected, and interactive lobe segmentation, average processing times were below 10minutes each and the mean overlaps were 88.1%, 91.8%, and 92.3%, respectively.