Automatic 3D Registration of Trabecular Bone Images Using a Collection of Regional 2D Registrations
Vasilic B, Lin W, Wehrli F, Magland J
University of Pennsylvania Medical Center
An efficient algorithm is presented for automatically registering serial high-resolution 3D MR images of trabecular bone (TB). Such a 3D registration is important in longitudinal studies which seek to identify small changes in trabecular architecture over time. One difficulty is that the soft tissue adjacent to bone is not rigidly connected to the TB region of interest (ROI). Thus, a successful technique must isolate and register only the ROI. An automatic segmentation algorithm is first used to roughly locate the region of interest within the baseline and follow-up scans. A collection of random 2D regions within the baseline ROI are then matched to the follow-up scan using the local trabecular pattern, and these local registrations are fit to a global 3D rigid transformation. The algorithm was tested on nine subjects who were scanned at two time points as a part of a clinical study. In each case, the fully automatic algorithm succeeded in less than 30 seconds with sub-pixel accuracy.