A Unified Framework for Atlas Based Brain Image Segmentation and Registration

We propose a unified framework in which atlas-based segmentation and non-rigid registration of the atlas and the study image are iteratively solved within a maximum-likelihood expectation maximization (ML-EM) algorithm. Both segmentation and registration processes minimize the same functional, i.e. the log-likelihood, with respect to classification parameters and the spatial transformation. We demonstrate how both processes can be integrated in a mathematically sound and elegant way and which advantages this implies for both segmentation and registration performance. This method (Extended EM, EEM) is evaluated for atlas-based segmentation of MR brain images on real data and compared to the standard EM segmentation algorithm without embedded registration component initialized with an affine registered atlas or after registering the atlas using a mutual information based non-rigid registration algorithm (II).