Non-rigid image registration by neural computations

This paper describes a neural computation based nonrigid registration methodology using multiple rigid transforms, in a piece-wise fashion, to model the registration process between images in a sequence. The registration methodology is a hybrid approach that combines registration without exact point correspondence via multi-object principal axes, and registration with point correspondence via polynomial transform. Neural computation is used to combine the derived individual principal axes solutions for each object in a committee machine formulation and to obtain the polynomial transform based on extracted control points using a multi layer perceptrons (MLP). Three examples are presented to demonstrate the techniques involved in the process. The first example uses four Gaussian clusters and focuses on the combination of the multiple transforms into a composite transform using finite mixture modeling techniques. The next examples present the complete process for prostate cancer registration and breast sequence analysis respectively. To verify performance, the results are compared to non-neural based implementations and other existing registration methods.