Adapting registration-based-segmentation for efficient segmentation of thoracic 4D MRI

Registration-based-segmentation is an accurate technique to segment target structures for thoracic 4D (3D + time) MRI data series that comprises a number of 3D MRI volumes acquired over several respiratory phases. However, directly applying registration-based segmentation techniques to segment the whole 4D MRI set will be inefficient. A reason for this inefficiency is that the tolerance number to terminate registration is usually set as a fixed value that can potentially lead the registration to exceed the point beyond what is required. This will result in unnecessary computational amount. In this study, we investigate the relationship between the optimal tolerance number and image similarity and proposed a manner that is based on spatio-temporal information to adaptive adjust registration tolerance.