Multimodal MR image registration using weakly supervised constrained affine network

Multimodal image registration is an important technique for many clinical applications. However, it is particularly challenging to obtain good spatial alignment. This paper introduces a novel architecture named the constrained affine network, which combines deformable image registration with affine transformation for multimodal MR image registration. A weakly supervised manner is adapted to train the network and anatomical labels are used in training. The network directly learns to predict a displacement vector field (DVF) between pairs of input images. Different from the existing deformable image registration methods based on the convolutional neural network (CNN), the method proposes a global constrained affine module, which can predict an affine transformation by pre-computing the range of affine parameters, and the model can be combined with a deformable registration network. We evaluated the proposed method on 3D multimodal medical images. Experimental results indicate that the proposed method has better performance.