A Diffeomorphic Approach to Multimodal Registration with Mutual Information: Applications to CLARITY Mouse Brain Images

Large Deformation Diffeomorphic Metric Mapping (LDDMM) is a widely used deformable registration algorithm for computing smooth invertible maps between various types of anatomical shapes such as landmarks, curves, surfaces or images. In this work, we specifically focus on the case of images and adopt an optimal control point of view so as to extend the original LDDMM with Sum of Squared Differences (SSD) matching term to a framework more robust to intensity variations, which is critical for cross-modality registration. We implement a mutual information based LDDMM (MI-LDDMM) algorithm and demonstrate its superiority to SSD-LDDMM in aligning 2D phantoms with differing intensity profiles. This algorithm is then used to register CLARITY mouse brain images to a standard mouse atlas despite their differences in grayscale values. We complement the approach by showing how a cascaded multi-scale method improves the optimization while reducing the run time of the algorithm.