Robust Unsupervised Image to Template Registration Without Image Similarity Loss

Slim Hachicha, Célia Le, Valentine Wargnier-Dauchelle, Michaël Sdika

Abstract: Although a giant step forward has been made in medical images analysis thanks to deep learning, good results still require a lot of tedious and costly annotations. For image registration, unsupervised methods usually consider the training of a network using classical registration dissimilarity metrics. In this paper, we focus on the case of affine registration and show that this approach is not robust when the transform to estimate is large. We propose an unsupervised method for the training of an affine image registration network without using dissimilarity metrics and show that we are able to robustly register images even when the field of view is significantly different in the image.