Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TLRN: Temporal Latent Residual Networks For Large Deformation Image Registration

Nian Wu, Jiarui Xing, Miaomiao Zhang

TL;DR

The paper addresses large, time-varying deformations in time-series image registration, notably in cardiac cine MRI. It introduces the Temporal Latent Residual Network (TLRN), which learns latent velocity fields in a temporal latent space $\mathcal{Z}$ and refines deformation features over time via residual blocks. The model uses an unsupervised registration backbone (U-Net) to map image sequences to $\{z^1,\dots,z^T\}$ and back, with a temporal residual learning module enforcing temporal consistency. Experiments on synthetic 2D sequences and real cine-CMR data show improved registration accuracy, better regularization of deformation fields, and reduced incidence of invalid deformations, with code publicly available.

Abstract

This paper presents a novel approach, termed {\em Temporal Latent Residual Network (TLRN)}, to predict a sequence of deformation fields in time-series image registration. The challenge of registering time-series images often lies in the occurrence of large motions, especially when images differ significantly from a reference (e.g., the start of a cardiac cycle compared to the peak stretching phase). To achieve accurate and robust registration results, we leverage the nature of motion continuity and exploit the temporal smoothness in consecutive image frames. Our proposed TLRN highlights a temporal residual network with residual blocks carefully designed in latent deformation spaces, which are parameterized by time-sequential initial velocity fields. We treat a sequence of residual blocks over time as a dynamic training system, where each block is designed to learn the residual function between desired deformation features and current input accumulated from previous time frames. We validate the effectivenss of TLRN on both synthetic data and real-world cine cardiac magnetic resonance (CMR) image videos. Our experimental results shows that TLRN is able to achieve substantially improved registration accuracy compared to the state-of-the-art. Our code is publicly available at https://github.com/nellie689/TLRN.

TLRN: Temporal Latent Residual Networks For Large Deformation Image Registration

TL;DR

The paper addresses large, time-varying deformations in time-series image registration, notably in cardiac cine MRI. It introduces the Temporal Latent Residual Network (TLRN), which learns latent velocity fields in a temporal latent space and refines deformation features over time via residual blocks. The model uses an unsupervised registration backbone (U-Net) to map image sequences to and back, with a temporal residual learning module enforcing temporal consistency. Experiments on synthetic 2D sequences and real cine-CMR data show improved registration accuracy, better regularization of deformation fields, and reduced incidence of invalid deformations, with code publicly available.

Abstract

This paper presents a novel approach, termed {\em Temporal Latent Residual Network (TLRN)}, to predict a sequence of deformation fields in time-series image registration. The challenge of registering time-series images often lies in the occurrence of large motions, especially when images differ significantly from a reference (e.g., the start of a cardiac cycle compared to the peak stretching phase). To achieve accurate and robust registration results, we leverage the nature of motion continuity and exploit the temporal smoothness in consecutive image frames. Our proposed TLRN highlights a temporal residual network with residual blocks carefully designed in latent deformation spaces, which are parameterized by time-sequential initial velocity fields. We treat a sequence of residual blocks over time as a dynamic training system, where each block is designed to learn the residual function between desired deformation features and current input accumulated from previous time frames. We validate the effectivenss of TLRN on both synthetic data and real-world cine cardiac magnetic resonance (CMR) image videos. Our experimental results shows that TLRN is able to achieve substantially improved registration accuracy compared to the state-of-the-art. Our code is publicly available at https://github.com/nellie689/TLRN.
Paper Structure (11 sections, 4 equations, 5 figures)

This paper contains 11 sections, 4 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Background: Diffeomorphic Image Registration
  3. Our Method: TLRN
  4. Network Architecture
  5. Experimental Evaluation
  6. Datasets
  7. Experimental Design
  8. Results
  9. Conclusions & Discussion
  10. Acknowledgments.
  11. Disclosure of Interests.

Figures (5)

  • Figure 1: An overview of our proposed network TLRN.
  • Figure 2: Left to right: examples of deformed reference/source image across time. Top to bottom: a comparison of deformed images and transformation fields predicted from our model TRLN and baselines.
  • Figure 3: A comparison of MSE between deformed and target time-series images.
  • Figure 4: Left to right: examples of CMR image videos and overlaid LV myocardium segmentation maps. Top to bottom: a comparison of manually delineated segmentation lables vs. propagated segmentation from all methods.
  • Figure 5: Left to right: a comparison of dice vs. Hausdorff distance score on predicted LV myocardium segmentation labels from all methods over the time frames $\tau (1 \sim 6)$.