Approximate nearest neighbors: towards removing the curse of dimensionality
STOC '98 Proceedings of the thirtieth annual ACM symposium on Theory of computing
Multidimensional binary search trees used for associative searching
Communications of the ACM
Registration of Challenging Image Pairs: Initialization, Estimation, and Decision
IEEE Transactions on Pattern Analysis and Machine Intelligence
| Hi-index | 0.00 |
Accurate affine registrations are crucial for many applications in medical image analysis. Within the Osteoarthritis Initiative (OAI) dataset we have observed a failure rate of approximately 4% for direct affine registrations of knee MRI without manual initialisation. Despite this, the problem of robust affine registration has not received much attention in recent years. With the increase in large medical image datasets, manual intervention is not a suitable solution to achieve successful affine registrations. We introduce a framework to improve the robustness of affine registrations without prior manual initialisations. We use 10,307 MR images from the large dataset available from the OAI to model the low dimensional manifold of the population of unregistered knee MRIs as a sparse k-nearest-neighbour graph. Affine registrations are computed in advance for nearest neighbours only. When a pairwise image registration is required the shortest path across the graph is extracted to find a geodesic path on the empirical manifold. The precomputed affine transformations on this path are composed to find an estimated transformation. Finally a refinement step is used to further improve registration accuracy. Failure rates of geodesic affine registrations reduce to 0.86% with the registration framework proposed.