A 3D+Time Spatio-temporal Model for Joint Segmentation and Registration of Sparse Cardiac Cine MR Image Stacks

We previously developed a hybrid spatio-temporal method for the segmentation of the left ventricle in 2D+time magnetic resonance (MR) image sequences and here extend this model-based approach towards 3D+time sparse stacks of cine MR images with random orientation. The presented method combines an explicit landmark based statistical geometric model of the inter-subject variability at the end-diastolic and end-systolic time frames with an implicit geometric model that constraints the intra-subject frame-to-frame temporal deformations through deterministic non-rigid image registration of adjacent frames. This hybrid model is driven by both local and global intensity similarity, resulting in a combined spatio-temporal segmentation and registration approach. The advantage of our hybrid model is that the segmentation of all image slices and of the whole sequence can be performed at once, guided by shape and intensity information of all time frames. In addition, prior shape and intensity knowledge are incorporated in order to cope with ambiguity in the images, while keeping training requirements limited.