Image registration and segmentation in longitudinal MRI using temporal appearance modeling

With increasing use of subject-specific longitudinal imaging for assessment of development, degeneration and disease progression, there is a clear need for image analysis segmentation/registration tools dedicated to 4D image time series. Previous work has mostly focused on temporal modeling of geometric deformations and shape changes, assuming that image intensity changes can be normalized. However, in studies of early infant development or aging, e.g., we encounter low contrast and appearance alterations due to tissue property changes which pose challenges to temporal registration and 4D segmentation. The two problems are linked since registration can be solved if appearance changes are accounted for, but 4D segmentation requires registration of image time series. In this paper, we propose to integrate a temporal appearance change model into diffeomorphic registration thus accounting for such variations, where voxel-wise intensity model parameters are calculated jointly with temporal image coregistration. Moreover, we demonstrate novel 4D segmentation of co-registered images that uses local intensity change rather than intensity itself via Gaussian mixture model. Both methods can be seen as two stages of an integrated registration/segmentation framework for 4D time-discrete image data making use of the same underlying model of longitudinal appearance changes. We demonstrate feasibility of the new approach with verification on longitudinal, multimodal pediatric MRI of infants in the age range neonates to 24 months.