Automatic non-rigid mismatch correction algorithm for CT-CT longitudinal oncology studies

Significant coverage mismatches, non-rigid motion, structural variability of anatomy, and intensity variations in contrast studies make automatic correction for multi-time point CT oncology studies a difficult problem. Also, since multiple volumes have to be non-rigidly co-aligned, it is important and challenging to maintain consistency of maps during non-rigid registration [NRR]. We present a fully automatic algorithm to handle above challenges using, 1) A novel graph-matching approach to initialize rigid registration under large coverage mismatches 2) An approximate, 1D diffeomorphic framework using patch based normalized image gradients to handle NRR. The above combination makes our method robust, fast (≈ 6 s for matching a pair of volumes on standard CPUs for high res. CT cases) and can be used to handle multiple volumes seamlessly. Validation of the proposed algorithm is shown on 17 pairwise NRR experiments on challenging multi-time point oncology studies.