Fast and accurate global geodesic registrations using knee MRI from the Osteoarthritis Initiative

Registration is important for many applications in medical image analysis. Affine registration of knee MR images can suffer failures due to large anatomical and articulated pose variations. This work is demonstrated using 2743 MR images from the Osteoarthritis Initiative (OAI) public access dataset. With such large datasets any manual interventions to aid registration success are not feasible and so full automation with high accuracy is of paramount importance. Additionally, computing exhaustive pairwise registrations across the OAI dataset is very computationally expensive. We present a sparse geodesic registration method that increases accuracy of pairwise registration and also enables fast online computation of registration. We then propose two novel methods to reduce registration error over the graph. Firstly we use all precomputed transformations to infer transformation errors for each edge, through assuming global registration cycle consistency across a sparse graph. In conjunction with this, we suggest fusing multiple successful registrations as a strategy to mitigate small errors in each transformation in the graph. It is shown that, in combination, these techniques achieve more accurate pairwise registration results than both geodesic registration and direct pairwise registration. This paper addresses accuracy of registrations, speed of online computation and is demonstrated on a large scale dataset.