Fully automatic segmentation of pulmonary segments from computed tomography chest scans

E. van Rikxoort, H. Gietema, M. Prokop and B. van Ginneken

Annual Meeting of the Radiological Society of North America 2006.

PURPOSE: The lung lobes are subdivided into pulmonary segments. There are 10 segments in the right lung and 9 in the left lung. An algorithm is presented for automatic identification of pulmonary segments from computed tomography scans. METHOD AND MATERIALS: Data was taken from the Nelson study, a lung cancer screening program with low dose CT (Philips Mx8000IDT, 16 x 0.75 mm collimation, 30 mAs). In this study, all abnormal findings are annotated by expert radiologists and for each finding the pulmonary segment in which it resides is recorded. From all scans with more than 5 findings, 200 scans were randomly selected. From these, 190 scans with 1690 findings were used for training the system; the remaining 10 scans with 196 findings were used for evaluation. In these evaluation scans, a second expert recorded the pulmonary segments for all findings. Two automatic algorithms were developed to find, respectively, the lungs and the major and minor pulmonary fissures within the lungs. From these segmentations, a number of features can be assigned to each lung voxel based on its relative position in the lung, and its distance from and position relative to the fissures. Using the known segment labels for all findings in the training data, a k-nearest neighbor classifier is trained that can assign each lung voxel in an unseen scan to a pulmonary segment. RESULTS: The second expert agreed with the original recordings in 134 of the 196 findings (68%). The accuracy of the automatic system in assigning pulmonary segment labels evaluated on all 196 findings was 77%. This increased to 85% using the set of 134 finding for which there was consensus among experts. When the system made an error, a location was always assigned to a neighboring segment. CONCLUSION: A position in the lungs can be automatically assigned to a pulmonary segment with an accuracy comparable to that of human observers. CLINICAL RELEVANCE/APPLICATION: Automatic segmentation of pulmonary segments is an important feature for clinical workstations for lung analysis and for a range of applications in computer aided diagnosis