Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Encoding Matching Criteria for Cross-domain Deformable Image Registration

Zhuoyuan Wang, Haiqiao Wang, Yi Wang

TL;DR

This work tackles cross-domain deformable image registration by introducing a registration-oriented framework that explicitly models matching criteria through two specialized encoders: Encoder-G for general medical features and Encoder-S augmented with a Structural Embedding Module to capture self-similarity. A ModeT-based decoder fuses pyramid features from both encoders to estimate a deformation field $\phi$ that aligns moving and fixed images, with a three-stage training scheme and one-shot learning for efficient domain adaptation. On three MRI datasets, the method achieves strong single-domain accuracy and superior cross-domain generalization, significantly reducing adaptation time compared with traditional optimization-based methods. The approach offers practical impact for rapid, cross-domain registration in clinical workflows, enabling reuse across scanners and modalities with minimal labeled data and computation.

Abstract

Most existing deep learning-based registration methods are trained on single-type images to address same-domain tasks.However, cross-domain deformable registration remains challenging.We argue that the tailor-made matching criteria in traditional registration methods is one of the main reason they are applicable in different domains.Motivated by this, we devise a registration-oriented encoder to model the matching criteria of image features and structural features, which is beneficial to boost registration accuracy and adaptability.Specifically, a general feature encoder (Encoder-G) is proposed to capture comprehensive medical image features, while a structural feature encoder (Encoder-S) is designed to encode the structural self-similarity into the global representation.Extensive experiments on images from three different domains prove the efficacy of the proposed method. Moreover, by updating Encoder-S using one-shot learning, our method can effectively adapt to different domains.The code is publicly available at https://github.com/JuliusWang-7/EncoderReg.

Encoding Matching Criteria for Cross-domain Deformable Image Registration

TL;DR

This work tackles cross-domain deformable image registration by introducing a registration-oriented framework that explicitly models matching criteria through two specialized encoders: Encoder-G for general medical features and Encoder-S augmented with a Structural Embedding Module to capture self-similarity. A ModeT-based decoder fuses pyramid features from both encoders to estimate a deformation field that aligns moving and fixed images, with a three-stage training scheme and one-shot learning for efficient domain adaptation. On three MRI datasets, the method achieves strong single-domain accuracy and superior cross-domain generalization, significantly reducing adaptation time compared with traditional optimization-based methods. The approach offers practical impact for rapid, cross-domain registration in clinical workflows, enabling reuse across scanners and modalities with minimal labeled data and computation.

Abstract

Most existing deep learning-based registration methods are trained on single-type images to address same-domain tasks.However, cross-domain deformable registration remains challenging.We argue that the tailor-made matching criteria in traditional registration methods is one of the main reason they are applicable in different domains.Motivated by this, we devise a registration-oriented encoder to model the matching criteria of image features and structural features, which is beneficial to boost registration accuracy and adaptability.Specifically, a general feature encoder (Encoder-G) is proposed to capture comprehensive medical image features, while a structural feature encoder (Encoder-S) is designed to encode the structural self-similarity into the global representation.Extensive experiments on images from three different domains prove the efficacy of the proposed method. Moreover, by updating Encoder-S using one-shot learning, our method can effectively adapt to different domains.The code is publicly available at https://github.com/JuliusWang-7/EncoderReg.
Paper Structure (14 sections, 4 equations, 6 figures, 1 table)

This paper contains 14 sections, 4 equations, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Network Overview
  4. General Feature Encoder
  5. Structural Feature Encoder
  6. Training Strategy
  7. One-shot Learning for Inference Stage
  8. Experiments
  9. Datasets.
  10. Implementation Details.
  11. Comparison Methods and Evaluation Metrics.
  12. Single-Domain Results.
  13. Cross-Domain Results.
  14. Conclusion

Figures (6)

  • Figure 1: The registration for domain A (abdomen MRI) and domain B (brain MRI). Model_A and Model_B represent the deep learning-based registration networks trained on domain A and domain B, respectively. It can be observed that Model_A and Model_B cannot provide accurate results for the cross-domain registration task, while the traditional SyN syn achieves satisfactory results.
  • Figure 2: The proposed registration network consists of the registration-oriented Encoder-G and Encoder-S, and the motion decomposition Transformer (ModeT) modet decoder.
  • Figure 3: General feature encoder (Encoder-G).
  • Figure 4: Structural feature encoder (Encoder-S) and the structural embedding module (SEM). $h$, $w$, $l$ and $c$ represent the height, width, length and channel number of the features, respectively. $N(x)$ denotes the $n \times n \times n$ neighborhood for voxel $x$.
  • Figure 5: The registration results on different datasets. The colored regions indicate the masks of different anatomical structures.
  • ...and 1 more figures