Registration

Landmark Detection in the Chest and Registration of Lung Surfaces in Computed Tomography Scans

Margrit Betke, Harrison Hong, Chekema Prince, and Jane P. Ko

We developed an automated system for registering chest CT images temporally. Our system detects anatomical landmarks in two CT scans and matches them to obtain an initial alignment of the chests. The lung surfaces are then segmented from the chests. We developed an efficient algorithm to establish correspondences of lung surface points. With this algorithm, lung surfaces are registered in an iterative closest-point process, improving the initial alignment step by step. We present a validation study that is based on registering vessel branch points within the lungs. We applied our method to align the lung surfaces of 10 pairs of chest CT scans and report a promising registration performance.


On the left, generic template images of the sternum, trachea, and vertebra. Next, a coronal view of a chest CT scan. The yellow line marks the most cranial image with visible lung (A), the purple line the axial image at the carina (B). The trachea in image A (green) is detected by correlation-based template matching using the generic trachea template on the left. Sternum and vertebra in image B (light and dark blue) are detected using their respective generic templates. The trachea in image B (green) is found using a template cropped online from the preceding axial image.
Initial landmark registration: Four points used for registration are shown for each scan: the center of the trachea cross-section in slice A and the centers of the cross-sections of sternum, trachea, and vertebra in slice B in each study. The landmarks in study 2 (green) are then be matched to the landmarks in study 1 (red).
Registration results for high-resolution lung surfaces. The lung surfaces are shown before and after registration. Zoomed-in views of the lungs are given on the right. The lungs in scan 1 are shown in red and in scan 2 prior to registration in green and after registration in blue. The registration process shifted the surfaces in scan 2 to the left and slightly rotated them to align with the surfaces in scan 1
Top views of the right lung are given before any processing (left), after the initial surface registration based on the landmark registration parameters (middle), and after 25 iterations of the lung surface registration (right). The surface in scan 1 is shown in red. The surface in scan 2 is shown in green prior to registration and in blue after registration.
Validation of surface registration. The radiologist established point correspondences of 42 vessel branching points in two CT studies of the same patient. On top, a coronal view of these points in study 1 (red) and study 2 (green) before registration. On the left, the points in study 2 (blue) are aligned to the points in study 1 (red) by the gold-standard rigid-body transformation that minimizes the sum of squared differences (SSD) between the 42 point pairs. On the right, the points in study 2 (blue) are aligned to the points in study 1 (red) using the rigid-body transformation that minimizes the SSD between corresponding point pairs on the lung surfaces using our registration algorithm. The translational difference between the two transformations is less than 5 mm. The differences in Euler angles are also small (2.1, 1.3, and 0.5 degrees).
Margrit Betke, Assistant Professor
Computer Science Department
Boston University
Email: betke@cs.bu.edu
URL: http://www.cs.bu.edu/faculty/betke
Phone: 617-353-8919
Fax: 617-353-6457