Reconstruction of blood vessel networks from a few perspective projections

We are investigating systems that accept a few two dimensional images that are perspective projections of blood vessels to reconstruct a three dimensional model of those vessels. This task is impossible unless a priori information is used; how this information is represented is widely regarded as a key issue. We describe the form that our system uses and explain why it is an improvement on previous representations. In particular, we show that the representation is extensible in that new information can be added to it at any time, and that the representation is task independent, in the sense that it can be used in many ways. We demonstrate its application to the problem of reconstruction and discuss how the representation can be "learned from observation".