A neural network approach for 3D surface modeling and registration

Surface based registration is commonly used in image aided surgery. This technique is extremely computationally expensive due to (1) the number of iterations required to search through the large parameter space and (2) the heavy computational load needed for determining the cost function (the distance between two surfaces). This is the main obstacle in pushing surface based registration for image guided surgery, where near real time registration is needed. Most attempts to reduce the computational burden, e.g., gradient descent and ICP, have been targeted at reducing the number of iterations for the optimization. In this paper, we propose to use a neural network to model the surface of the reference structure. This not only provides an accurate model for the surface but also a fast method for computing the cost function. For CT-CT spine registration, the time taken to register two spine surfaces is about 10 times faster compared to the commonly used triangular mesh modeling with similar registration accuracy.