Identification of lobar fissures in pathological lungs

Identification of lobar fissures in human lungs is a non-trivial task due to their variable shape and appearance, along with the low contrast and high noise in computed tomographic (CT) images. Pathologies in the lungs can further complicate this identification by deforming and/or disrupting the lobar fissures. Current algorithms rely on the general anatomy of the lungs to find fissures affected by pathologies. This can be unreliable as deformations and/or disruptions of these fissures will alter the general lung anatomy. To overcome this, we developed an algorithm with the following novelties: (1) a new application of neural network based texture analysis to generalize fissure regions; and (2) a new method of fissure surface identification. We tested our algorithm on CT image stacks from 8 anonymous patients with pathological lungs. Compared to manually segmented fissures, our algorithm produced an average mean difference of 0.71 mm and 0.68 mm for identifying the left and right oblique fissures, respectively. Using a 3-mm percentile measure, the algorithm yielded an average accuracy of 86.8% for the left oblique fissure with a mean worst-case error of 3.18 mm. For the right oblique fissure, the algorithm produced an accuracy of 88.8% with a mean worst-case error of 3.13 mm. The above results show feasibility of using our algorithm for identifying fissures in pathological lungs.