Joint registration and averaging of multiple 3D anatomical surface models

In biological imaging, three-dimensional (3D) reconstruction from serial sections is a powerful technique that produces graphical surface models of anatomical structures. Since any model is built from data gathered on a single animal, it is not statistically representative. To overcome this limitation, multiple 3D models must be registered and averaged. This paper describes a method that jointly registers and averages several 3D surface models. The algorithm iterates series of pairwise registrations between each model and the average of the remaining ones. The method has been tested on synthetic data and several sets of real neuroanatomical models. The results illustrate that the method is robust and converges rapidly. In practice, reconstructed 3D models generally contain the envelopes of several anatomical structures. Therefore, a generalized version of the algorithm is proposed that produces the average surface of each structure and globally registers all the structures. It is shown that the registration is more accurate when computed from the envelopes of several objects. Average 3D surface models upon which shape variability can be represented are obtained.