End-to-End Unsupervised Deformable Image Registration with a Convolutional Neural Network

PDF

In this work we propose a deep learning network for deformable image registration (DIRNet). The DIRNet consists of a convolutional neural network (ConvNet) regressor, a spatial transformer, and a resampler. The ConvNet analyzes a pair of fixed and moving images and outputs parameters for the spatial transformer, which generates the displacement vector field that enables the resampler to warp the moving image to the fixed image. The DIRNet is trained end-to-end by unsupervised optimization of a similarity metric between input image pairs. A trained DIRNet can be applied to perform registration on unseen image pairs in one pass, thus non-iteratively. Evaluation was performed with registration of images of handwritten digits (MNIST) and cardiac cine MR scans (Sunnybrook Cardiac Data). The results demonstrate that registration with DIRNet is as accurate as a conventional deformable image registration method with substantially shorter execution times.
Submitted 20 Apr 2017 to Computer Vision and Pattern Recognition [cs.CV]
Published 21 Apr 2017
http://arxiv.org/abs/1704.06065
http://arxiv.org/pdf/1704.06065.pdf

0 comments