3D reconstruction of tree structures from biplane pictures

This paper describes a system for three-dimensional reconstruction of tree-like objects from biplane pictures. Corresponding to medical X-ray angiography a wire-frame phantom of the human coronary tree is built. Pictures of the phantom are taken from two different views with 90/spl deg/ rotation. Main features of the tree structure are extracted and feature points are combined to segments. A subset of feature points is selected for correspondence finding and 3D reconstruction. The correspondence finding problem is formulated as a cost function and mapped onto a two dimensional binary Hopfield neural network. The cost function takes into account geometric constraints due to the imaging aperture. Results found by the neural network are close to matching results attained interactively.