Three dimensional image registration using artificial neural networks

Registration of three-dimensional medical images is important for correlation of images from different modalities and to be able to follow progression or regression of disease. In this paper, the authors investigate the use of artificial neural networks in registering simulated 3-D images. Backpropagation networks with 0 or 1 hidden layers accurately map between coordinate spaces which are rotated, translated, and linearly scaled in 3 dimensions. Mapping between coordinate spaces which are nonlinear related is less accurate. Functional link net type architecture and larger training sets appear to improve the accuracy on these non-linear mappings