Computer Vision News - January 2017

34 Computer Vision News Research Research 4. Airway shape features - PCA was used to overcome the variation between airways in the dataset due to size, position and rotation which stem from patient age and scan position and are not indicative of airway pathology. The PCA input for each object is a 3n dimensional stacked vector of mesh vertices where n is the number of vertices in the mesh; roughly n≈1700, and the output is 11 eigenvectors representing 90-99.5% of the variance. About Test Time - The LA-PDM detects airway deformation related to pediatric TB using the following computer-aided diagnostic pipeline: a chest CT of a pediatric patient with suspected TB is acquired, the airway is automatically segmented, the airway centerline and branch points are identified and landmarks are placed on the airway surface using the methods described in step 1 and 2 above. The airway template mesh for each 3-branch region (Trachea– RMB–LMB and RMB-RUL-BI – see figure below) is then registered to the segmented airway surface to develop vertex to vertex correspondence with the training set (step 3). The PDM, created using the training set, is then used to project the mesh vertices of the test case, and the first 11 modes of variation are extracted as features (step 4). A classifier, trained on features from the training set, is then used to classify each 3-branch region of the airways. This allows automatic detection of airway abnormality in each 3-branch region based on variation of the shape of each bronchi as well as variation with respect to neighboring bronchi.