A Deep Discontinuity-Preserving Image Registration Network

Abstract

Image registration aims to establish spatial correspondence across pairs, or groups of images, and is a cornerstone of medical image computing and computer-assisted-interventions. Currently, most deep learning-based registration methods assume that the desired deformation fields are globally smooth and continuous, which is not always valid for real-world scenarios, especially in medical image registration (e.g. cardiac imaging and abdominal imaging). Such a global constraint can lead to artefacts and increased errors at discontinuous tissue interfaces. To tackle this issue, we propose a weakly-supervised Deep Discontinuity-preserving Image Registration network (DDIR), to obtain better registration performance and realistic deformation fields. We demonstrate that our method achieves significant improvements in registration accuracy and predicts more realistic deformations, in registration experiments on cardiac magnetic resonance (MR) images from UK Biobank Imaging Study (UKBB), than state-of-the-art approaches.

Evaluation of image registration for measuring deformation fields in soft tissue mechanics

“…[53] Nevertheless, discontinuities might also occur during a patient screening where the movement of organs leads to local discontinuities and thus they should inspire further research in the field of image registration. This problem was recently investigated by using deep learning [54] to obtain better spatial transformation of two image sets without the need for globally smooth and continuous transformation fields. Although incorporation of machine learning is very interesting, the complexity could limit the spread among more users.…”

A segmentation-informed deep learning framework to register dynamic two-dimensional magnetic resonance images of the vocal tract during speech

“…Registration and segmentation can be related tasks, and there is increasing evidence that including segmentation information during the training of a registration CNN results in more accurate motion estimates [37] , [42] , [43] , [44] , [45] , [46] , [47] , [48] , [49] , [50] , [51] . Inclusion of such information is typically achieved by including region-overlap-based terms such as the Dice coefficient (DSC) in the CNN loss function.…”

“…The first approach is to use segmentations to modify the appearance of the images, in order to optimise the images for the registration task [49] , [50] , [51] . In this approach, the images are modified before being used as inputs to the registration CNNs either by multiplying them by binary segmentations [49] , [50] or by using a fully convolutional image transformer network whose loss function includes a region-overlap-based term [51] . The second approach is to use segmentations as well as images as inputs to the registration CNN [43] .…”

“…Similarly to traditional deformable registration methods, currently the majority of deep-learning-based methods are designed to estimate smooth and continuous displacement fields. Two methods have been developed to estimate displacement fields with discontinuities [50] , [52] . [52] is designed to capture sliding motion only, while [50] is designed to capture cardiac motion and its suitability for capturing motion where organs come into contact and then separate from each other has not yet been investigated.…”

“…Two methods have been developed to estimate displacement fields with discontinuities [50] , [52] . [52] is designed to capture sliding motion only, while [50] is designed to capture cardiac motion and its suitability for capturing motion where organs come into contact and then separate from each other has not yet been investigated.…”