A PROBABILISTIC NON-RIGID REGISTRATION FRAMEWORK USING LOCAL NOISE ESTIMATES

Accurate inter-subject registration of magnetic resonance (MR) images of the human brain is required to allow meaningful comparisons across groups of subjects. Some anatomical structures can be very difficult to match and this can result in intensity based registration approaches inferring complex and implausible mappings in some regions. In this work, we propose a generic probabilistic framework for non-rigid registration with a spatially varying trade-off between image information and regularisation. This trade-off is based on local estimates of misalignment "noise", which effectively increases regularisation in regions which are difficult to register. We demonstrate that the proposed method infers smoother, more plausible and slightly more accurate mappings for inter-subject registration of MR images of the human brain.