DWD-net: Cascaded local and global deep learning network for brain MR registration

At present, non-rigid registration of adult brain magnetic resonance (MR) images has been widely applied. Both traditional registration methods and those based on deep learning can achieve preferable registration effect. However, low registration accuracy and poor registration effect may occur in local areas and complex anatomical structures. Therefore, dynamically and weakly supervised method is proposed in this paper, which can dynamically learn different local regions. In order to achieve the method, a local loss function is designed. The Dual Weakly and Dynamically supervised deep learning registration network (DWD-net) is proposed. The DWD-net is divided into two parts: Local net (Lnet) and Global net (Gnet). Lnet will dynamically align difficultly-aligned local regions by the proposed method and loss function. Gnet can align the details of the whole image. Experimental results are presented on the OASIS-1 dataset show that the proposed DWD-net realizes higher accuracy on both difficultly-aligned local regions and global regions than state-of-the-art registration methods.