The promising initial registration results described herein provide various directions of future work. To date, we have tested this registration algorithm only on MR images, primarily due the availability of this particular set of data. There is interest in acquiring more datasets of different modalities, including MR Angiogram, CT, and PET to further examine this registration technique by means of the Retrospective Registration Evaluation Project [12].
Another area of further investigation is to include additional statistical models to the current framework. Non-linear bias fields present in the MR data can cause mismatches in intensity histograms between the training and test images. Registration using a prior on joint intensity information can be sensitive to these differences. Thus, there is interest in integrating the statistical intensity correction work of Wells, et al. [10] into this registration technique to both provide a more reliable intensity correspondence between training and test data, and perhaps assist in the process of segmenting the anatomical structures in the images.
Additional information such as prior knowledge of the shapes or relative positions of internal structures may help in the registration process. Such information certainly aids in segmentation, by offering guidelines such as what structures are most likely to appear next to other structures. Given that segmentation and registration are related in that the knowledge of one greatly assists in the computation of the other, this would imply that the addition of these types of priors may also assist in registration.