Computer Vision News - January 2017

Computer Vision News Computer Vision News Research 33 Research not affect the topology, to extract a 1-voxel thick branching centerline. Then, the method iteratively identifies branch points and labels each branch using the voxel connectivity of the branching centerline, starting from the trachea. Branch points are identified when a voxel of the centerline has at least three neighboring voxels, each coroneted to 26 neighbors (it's a 3d image, thus each voxel has at most 26 neighbors). Finally, the segmentation is labelled by using a distance transform from the labelled centerline. 2. Surface point projection - A mesh of the airway surface is constructed. Landmarks are needed to align the airways in the dataset in order to develop the PDM. However, airways do not have any distinct landmarks except the bifurcation points. Pseudo-landmarks were constructed: the centerline was smoothed and resampled equidistantly and pseudo-landmark points were projected onto the airway surface mesh orthogonally to the centerline at each resampled point (see figure below). This representation allows a set of corresponding surface points to be mapped onto each branch of each segmented airway. 3. Thin-plate-spline warp - One airway from the control dataset was selected as a template; A thin-plate-spline (TPS) warp was applied to align the template with each airway in the dataset so that each airway is represented with a corresponding mesh. TPS is an interpolation method useful for non-rigid registration – a surface with a set of landmarks is warped to the corresponding set of landmarks in a physically realistic way (see the following figure). Airway analysis from CT scans can help determine abnormal compression or deformation of the airways and this is key to detecting disease and visualizing affected airway regions