Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CAR: Contrast-Agnostic Deformable Medical Image Registration with Contrast-Invariant Latent Regularization

Yinsong Wang, Siyi Du, Shaoming Zheng, Xinzhe Luo, Chen Qin

TL;DR

The paper tackles multi-contrast MRI registration by addressing contrast variability with a contrast-agnostic deformable registration framework. It introduces random convolution-based contrast augmentation to simulate arbitrary contrasts and contrast-invariant latent regularization to align latent representations across contrasts, enabling a single model to register unseen contrasts. The method uses a Siamese encoder–decoder architecture with mono-contrast LNCC loss and a contrast-regularization term, formulated through the total loss $L_{total}$. Experiments on brain and cardiac MRI demonstrate improved registration accuracy and deformation regularity over state-of-the-art methods, with strong generalization to unseen contrasts; the authors also provide ablations validating the proposed components, and code is released.

Abstract

Multi-contrast image registration is a challenging task due to the complex intensity relationships between different imaging contrasts. Conventional image registration methods are typically based on iterative optimizations for each input image pair, which is time-consuming and sensitive to contrast variations. While learning-based approaches are much faster during the inference stage, due to generalizability issues, they typically can only be applied to the fixed contrasts observed during the training stage. In this work, we propose a novel contrast-agnostic deformable image registration framework that can be generalized to arbitrary contrast images, without observing them during training. Particularly, we propose a random convolution-based contrast augmentation scheme, which simulates arbitrary contrasts of images over a single image contrast while preserving their inherent structural information. To ensure that the network can learn contrast-invariant representations for facilitating contrast-agnostic registration, we further introduce contrast-invariant latent regularization (CLR) that regularizes representation in latent space through a contrast invariance loss. Experiments show that CAR outperforms the baseline approaches regarding registration accuracy and also possesses better generalization ability to unseen imaging contrasts. Code is available at \url{https://github.com/Yinsong0510/CAR}.

CAR: Contrast-Agnostic Deformable Medical Image Registration with Contrast-Invariant Latent Regularization

TL;DR

The paper tackles multi-contrast MRI registration by addressing contrast variability with a contrast-agnostic deformable registration framework. It introduces random convolution-based contrast augmentation to simulate arbitrary contrasts and contrast-invariant latent regularization to align latent representations across contrasts, enabling a single model to register unseen contrasts. The method uses a Siamese encoder–decoder architecture with mono-contrast LNCC loss and a contrast-regularization term, formulated through the total loss . Experiments on brain and cardiac MRI demonstrate improved registration accuracy and deformation regularity over state-of-the-art methods, with strong generalization to unseen contrasts; the authors also provide ablations validating the proposed components, and code is released.

Abstract

Multi-contrast image registration is a challenging task due to the complex intensity relationships between different imaging contrasts. Conventional image registration methods are typically based on iterative optimizations for each input image pair, which is time-consuming and sensitive to contrast variations. While learning-based approaches are much faster during the inference stage, due to generalizability issues, they typically can only be applied to the fixed contrasts observed during the training stage. In this work, we propose a novel contrast-agnostic deformable image registration framework that can be generalized to arbitrary contrast images, without observing them during training. Particularly, we propose a random convolution-based contrast augmentation scheme, which simulates arbitrary contrasts of images over a single image contrast while preserving their inherent structural information. To ensure that the network can learn contrast-invariant representations for facilitating contrast-agnostic registration, we further introduce contrast-invariant latent regularization (CLR) that regularizes representation in latent space through a contrast invariance loss. Experiments show that CAR outperforms the baseline approaches regarding registration accuracy and also possesses better generalization ability to unseen imaging contrasts. Code is available at \url{https://github.com/Yinsong0510/CAR}.
Paper Structure (7 sections, 3 equations, 3 figures, 3 tables)

This paper contains 7 sections, 3 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Contrast Augmentation via Random Convolution
  4. Contrast-Invariant Latent Regularization
  5. Overall Network Architecture and Loss Function
  6. Experiments and Results
  7. Conclusion

Figures (3)

  • Figure 1: Overview of the contrast-agnostic registration framework. The input fixed and moving image pair is fed to two Siamese encoders to produce contrast-invariant latent representations, which are taken by the decoder to predict the deformation field.
  • Figure 2: Qualitative results on CamCAN dataset.
  • Figure 3: Qualitative results on CMRxRecon dataset using T1 Mapping data for training.