2D X-ray airway tree segmentation by 3D deformable model projection and registration

Chest x-rays (CXR) still play an important role in detection of airway disease. Airway stenosis is a key sign of disease such as paediatric tuberculosis but is challenging to measure in CXR because the airway has relatively low contrast compared to overlying structures. In this study a novel approach to identify airways in CXRs is introduced, using a 3D statistical shape model to guide the segmentation. The 3D model is projected onto the CXR and aligned to the airways using four manual landmarks. The 3D shape model is then fitted to each CXR using an energy function based on image gradient, anatomical shadow and a regularisation term. Anatomical shadow is a novel feature introduced to detect the airway even without a clearly defined boundary. The algorithm achieved a mean error of 6.8±2.6 pixels (0.82±0.31 mm) on the 31 patient test set, a 25± 17% improvement on the initial linear alignment.