Biomechanics-informed Neural Networks for Myocardial Motion Tracking in MRI

Abstract

Image registration is an ill-posed inverse problem which often requires regularisation on the solution space. In contrast to most of the current approaches which impose explicit regularisation terms such as smoothness, in this paper we propose a novel method that can implicitly learn biomechanics-informed regularisation. Such an approach can incorporate application-specific prior knowledge into deep learning based registration. Particularly, the proposed biomechanics-informed regularisation leverages a variational autoencoder (VAE) to learn a manifold for biomechanically plausible deformations and to implicitly capture their underlying properties via reconstructing biomechanical simulations. The learnt VAE regulariser then can be coupled with any deep learning based registration network to regularise the solution space to be biomechanically plausible. The proposed method is validated in the context of myocardial motion tracking on 2D stacks of cardiac MRI data from two different datasets. The results show that it can achieve better performance against other competing methods in terms of motion tracking accuracy and has the ability to learn biomechanical properties such as incompressibility and strains. The method has also been shown to have better generalisability to unseen domains compared with commonly used L2 regularisation schemes.

MICDIR: Multi-scale Inverse-consistent Deformable Image Registration using UNetMSS with Self-Constructing Graph Latent

“…But, owing to the fact that GANs are prone to generate unrealistic imagery and the difficulty around explaining them might not always be a preferred solution. However, there are some intuitive yet efficient ways of regularising these GANbased networks for producing realistic images, such as cycle consistency (Zhang, 2018;Kim et al, 2020), bio-mechanics informed regulariser (Qin et al, 2020), etc. However, the existing research either avoided considering or failed to incorporate information about structural connectivity inside the brain.…”

Double-Uncertainty Guided Spatial and Temporal Consistency Regularization Weighting for Learning-based Abdominal Registration

“…Regularization preserves the smoothness and penalizes unnecessary complexity of the estimated transformation, and hence reduces the solution space of image registration. For LBR, researchers have proposed task-specific regularization, e.g., population-level statistics [3] and biomechanical model [12,17], to replace traditional linear elasticity [15] and thin-plate spline bending energy [4]. Like traditional regularization, these methods all belong to the category of applying spatial constraints to the transformation.…”