An Efficient Airway Tree Segmentation Method Robust to Leakage Based on Shape Feature Optimization

Abstract

The main problem with most of 3D image segmentation methods which significantly influences their accuracy is the leakage occurred during the process of segmentation. In the case of airway tree segmentation, due to an imaging artifact or a thin airway wall, the contrast between the air and the airway wall can locally decrease which leads to allow the region-growing approach to move from the inside of the airway to the pulmonary parenchyma. This begins the leaks phenomena to build and large parts of the lungs can be erroneously marked as the airway tree. A user intervention is required in this case to detect the leakage point and restart the segmentation on the whole or part of the segmented area with new parameters. This makes the algorithm very exhaustive and more dependent on the user interaction.

The new strategy presented here is to prevent the leakage from its origination by taking the advantage of the fact that the airway branches are cylindrically shaped objects. This has been achieved by employing a mathematical shape optimization approach based on the radial gradient of voxels along the cylindrical axes to retain cylindrical properties of the airway branches during the process of segmentation. The proposed cost function consists of two parts named cylindricalshape feature extraction and error smoothness term. The first term approaches to its minimum when underlying voxels are arranged on a cylindrical shape. The role of the second term is to control and smooth the final error and simultaneously to overcome the local minima’s problem.

We first applied the cost function on the simulated cylindrical objects with spongy structure to model the leakage with holes and tunnel. The impact of each term on the final error and the convergence of the algorithm while approaching to its global minimum are evaluated.