CycleMorph: Cycle Consistent Unsupervised Deformable Image Registration

Boah Kim, Dong Hwan Kim, Seong Ho Park, Jieun Kim, June-Goo Lee, Jong Chul Ye

Introduction

Deformable image registration is a fundamental task in medical imaging, essential for applications such as longitudinal study of disease progression and atlas-based segmentation. While traditional variational methods offer high accuracy, their iterative nature makes them computationally prohibitive for real-time clinical use. Recent deep learning (DL) approaches have significantly accelerated this process, yet they frequently encounter a "degeneracy" problem: the generated deformation fields often lack diffeomorphic properties, leading to unphysical "folding" or tearing of anatomical structures.

CycleMorph addresses these challenges by introducing a cycle-consistent unsupervised learning framework. By enforcing a geometric "round-trip" constraint, the model ensures that transformations are reversible and topology-preserving, providing high-accuracy registrations at speeds orders of magnitude faster than conventional algorithms.

Methodology

Cycle Consistency Constraint

The core innovation of CycleMorph is the application of cycle consistency to the geometry of image registration. Unlike previous DL methods that might only enforce consistency on the deformation vectors, CycleMorph enforces it at the pixel level.

The CycleMorph architecture utilizes dual registration networks ($G_X, G_Y$) to learn forward and reverse transformations, ensuring geometric consistency through a cycle loop.

The system employs two convolutional neural networks (CNNs): $G_X$ for the forward deformation (moving image $X$ to target $Y$) and $G_Y$ for the reverse. The cycle loss measures how closely a "re-deformed" image—warped to the target and then warped back—matches the original source. This acts as an implicit regularizer that penalizes non-physical mappings and prevents topological folding without requiring the heavy computational overhead of explicit diffeomorphic integration layers during inference.

Multi-scale Implementation

To handle high-resolution 3D volumes, such as liver CT scans, while staying within GPU memory limits, the authors implemented a two-stage multi-scale strategy.

The multi-scale registration pipeline involves a global registration stage on downsampled volumes followed by a local patch-based registration to capture fine anatomical details.

In the training stage, a global registration model is first trained on downsampled volumes. Subsequently, a local registration model is trained on high-resolution patches. During the test stage, the global deformation field is fused with the local patch deformations to produce a final, high-fidelity registration field ($\phi_{final}$).

Experimental Setup and Datasets

The framework was validated across three distinct datasets to demonstrate its versatility:

  • Brain MRI (OASIS-3): 1,249 T1-weighted scans used for atlas-based registration.
  • Multiphase Liver CT: 555 scans featuring four different contrast-enhanced phases, testing the model's robustness to significant intensity and shape variations.
  • Facial Expression Image (RaFD): 1,608 2D images used to evaluate registration across varying facial expressions.

To handle varying intensities in contrast-enhanced CT, the model utilizes local cross-correlation (LCC) within its registration loss, ensuring stability even when source and target images exhibit different contrast characteristics.

Key Results and Discussion

Topology and Accuracy

CycleMorph demonstrates superior topological preservation compared to baselines like VoxelMorph. In brain MRI tasks, it significantly reduced the occurrence of non-positive Jacobian determinants (a primary indicator of folding). Furthermore, it achieved higher Dice scores across nearly all anatomical structures in the OASIS-3 dataset, indicating better alignment of complex boundaries.

Clinical Utility and Speed

In liver CT registration, CycleMorph reduced the Target Registration Error (TRE) and provided more reliable measurements of tumor size changes. The efficiency gains are stark: while traditional tools like Elastix require approximately 20 minutes for a 3D registration, CycleMorph's global registration takes just 4 seconds, with the complete multi-scale pipeline finishing in 41 seconds.

Conclusion

CycleMorph provides a robust, unsupervised solution for deformable image registration that bridges the gap between the speed of deep learning and the topological integrity required for medical applications. While the current two-stage training process for multi-scale registration adds some complexity, the resulting framework offers a highly practical tool for real-time clinical image analysis.

Summary updated

Abstract

Image registration is a fundamental task in medical image analysis. Recently, deep learning based image registration methods have been extensively investigated due to their excellent performance despite the ultra-fast computational time. However, the existing deep learning methods still have limitation in the preservation of original topology during the deformation with registration vector fields. To address this issues, here we present a cycle-consistent deformable image registration. The cycle consistency enhances image registration performance by providing an implicit regularization to preserve topology during the deformation. The proposed method is so flexible that can be applied for both 2D and 3D registration problems for various applications, and can be easily extended to multi-scale implementation to deal with the memory issues in large volume registration. Experimental results on various datasets from medical and non-medical applications demonstrate that the proposed method provides effective and accurate registration on diverse image pairs within a few seconds. Qualitative and quantitative evaluations on deformation fields also verify the effectiveness of the cycle consistency of the proposed method.
BibTeX
@article{kim2021cyclemorph,
  author = {Boah Kim and Dong Hwan Kim and Seong Ho Park and Jieun Kim and June-Goo Lee and Jong Chul Ye},
  title = {{CycleMorph: Cycle Consistent Unsupervised Deformable Image Registration}},
  journal = {Medical Image Analysis},
  year = {2021},
  url = {https://openreview.net/forum?id=4wLubvyBz7x},
}
Figure