A new method to validate thoracic CT-CT deformable image registration using auto-segmented 3D anatomical landmarks
Background. Deformable image registrations are prone to errors in aligning reliable anatomically features. Consequently, identification of registration inaccuracies is important. Particularly thoracic three-dimensional (3D) computed tomography (CT)-CT image registration is challenging due to lack of contrast in lung tissue. This study aims for validation of thoracic CT-CT image registration using auto-segmented anatomically landmarks.
Material and methods. Five lymphoma patients were CT scanned three times within a period of 18 months, with the initial CT defined as the reference scan. For each patient the two successive CT scans were registered to the reference CT using three different image registration algorithms (Demons, B-spline and Affine). The image registrations were evaluated using auto-segmented anatomical landmarks (bronchial branch points) and Dice Similarity Coefficients (DSC). Deviation of corresponding bronchial landmarks were used to quantify inaccuracies in respect of both misalignment and geometric location within lungs.
Results. The median bronchial branch point deviations were 1.6, 1.1 and 4.2 (mm) for the three tested algorithms (Demons, B-spline and Affine). The maximum deviations (> 15 mm) were found within both Demons and B-spline image registrations. In the upper part of the lungs the median deviation of 1.7 (mm) was significantly different (p < 0.02) relative to the median deviations of 2.0 (mm), found in the middle and lower parts of the lungs. The DSC revealed similar registration discrepancies among the three tested algorithms, with DSC values of 0.96, 0.97 and 0.91, for respectively Demons, B-spline and the Affine algorithms.
Conclusion. Bronchial branch points were found useful to validate thoracic CT-CT image registration. Bronchial branch points identified local registration errors > 15 mm in both Demons and B-spline deformable algorithms.