An Unsupervised Approach for 3D Medical Image Registration

Deformable image registration is of great importance in many clinical applications. In this work, we propose an unsupervised end-to-end registration method for 3D medical images. The proposed method takes a pair of moving and fixed images as input and directly estimates the spatial transformation parameters, which enable spatial transformation layer to generate the registered image. In particular, the registration network is designed in the underlying M-Net architecture that consists of encoding path, decoding path, left leg, and right leg. Moreover, we introduce a novel loss function to guide the training. The proposed method is evaluated on the public brain image dataset ADNI. Experimental results demonstrate that our method achieves promising performance.