3D lung registration using splineMIRIT and Robust Tree Registration (RTR)

Intra-patient registration of lung CT scans acquired at a different time points or inspiration levels is a valuable examination tool to study multiple lung images. It allows to study ventilation or other functional information of the lungs. In this paper, two 3D lung registration methods are presented. The first method, splineMIRIT, uses voxel based non rigid image registration. It is based on mutual information as similarity measure, a B-spline mesh to model the deformation and B-spline image interpolation. The second method, Robust Tree Registration (RTR), extends the first by including robust 3D registration of the vessel trees found in both images. The tree is represented by intrinsic matrices containing the geodesic or Euclidean distance between each pair of detected bifurcations. This representation is independent of the reference frame. Marginalization of point pair probabilities based on the intrinsic matrices provides soft assign correspondences between the two trees. This global correspondence model is combined with local bifurcation similarity models, based on the local gray value distribution. Finally, hard correspondences are deducted from the model. The correspondences between bifurcations are added to splineMIRIT as an additional similarity measure. The method is validated on the EMPIRE10 data set. Both algorithms perform well. Comparing splineMIRIT and RTR shows that on average the results slightly improve when the robust tree registration is added, leading to a 15 and 13 place, respectively, in the “Grand Challenges in Medical Image Analysis” workshop of MICCAI 2010.