Interactive lung segmentation in CT scans with severe abnormalities

摘要

Estimation of the volume of the lungs and the viable lung tissue is an important step in the management of patients with severe pulmonary disease. The presence of gross pathology makes it impossible to perform lung segmentation automatically and reliably in CT scans of such patients. An interactive system for lung segmentation is presented, based on precomputed compact regions with homogeneous texture for which general texture feature have been computed. A statistical classifier trained on prior data has classified these regions beforehand and the user corrects any errors until the segmentation of an entire slice is correct. The system proceeds to subsequent slices, which were preclassified using a combined classification strategy that uses both the prior data and the previously approved slices from the test scan. The resulting lung segmentations show a large overlap and a small average boundary distance when compared to completely manual delineations of the lung borders. The lung segmentations can then be used as input for a similar interactive system to determine the viable lung volume.