An Unsupervised Deep Learning Based Method For Fast Chest Image Registration

Image registration,the process of aligning two or more images,is a widely used core technique of image processing tasks.With the rapid development of science and technology,on the one hand,medical imaging technologies has been developed and clinical imaging devices have been popularized in society.This lead to the surging of large amounts of high dimensional images with rich structural and functional information,which becomes a big challenge to medical image registration.On the other hand,many kinds of image registration methods have been developed and applied in challenging areas and practical tasks.Although the performance of current methods is good,the computation cost of these methods is generally high.The high computation cost is mainly due to the iterative optimization problem of non-learning-based deformable registration.In this work we propose an unsupervised learning-based framework for deformable image registration.A convolutional neural network is used to learn a mapping from the input image pair to a dense displacement vector field that is then applied to the moving image by a spatial transformation function.The model is trained by optimizing a similarity metric between the warped moving image and the fixed image,with no need for any supervised information such as segmentation masks,manually annotations or synthetic transformations.We apply explicitly and implicitly spatial regularization terms to constrain the models to estimate smooth transformations for two kinds of image datasets,respectively.We evaluate the proposed model on a private dataset of chest radiographs and a publicly available dataset of chest CT images.The results demonstrate that the accuracy of the model is comparable to that of the conventional image registration while executing orders of magnitude faster.