Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Boosting Medical Image Synthesis via Registration-guided Consistency and Disentanglement Learning

Chuanpu Li, Zeli Chen, Yiwen Zhang, Liming Zhong, Wei Yang

TL;DR

This work tackles misalignment noise in medical image synthesis by integrating a registration-guided consistency framework with an anatomy consistency disentanglement synthetic module. It introduces a registration module (deformation generator $R_{\Phi}$ and resampler $R_S$) paired with an alignment loss $L_{align}$ to enforce geometry preservation across synthesis before and after applying the same deformation field, and an ACDS module that disentangles anatomical content from modality styles using dedicated encoders and generators, guided by adversarial, self-reconstruction, cycle-consistency, and anatomy-consistency losses. Experimental results on an in-house abdominal CECT-CT dataset and the SynthRAD2023 pelvic MR-CT dataset demonstrate superior quantitative performance (MAE, PSNR, SSIM) and clearer anatomical boundaries, with ablations confirming the contributions of both registration-guided consistency and ACDS. Overall, the approach provides a practical, geometry-preserving framework for robust cross-modality medical image synthesis, reducing misalignment artifacts and improving structural fidelity for clinical tasks such as MR-to-CT and CECT-to-CT synthesis.

Abstract

Medical image synthesis remains challenging due to misalignment noise during training. Existing methods have attempted to address this challenge by incorporating a registration-guided module. However, these methods tend to overlook the task-specific constraints on the synthetic and registration modules, which may cause the synthetic module to still generate spatially aligned images with misaligned target images during training, regardless of the registration module's function. Therefore, this paper proposes registration-guided consistency and incorporates disentanglement learning for medical image synthesis. The proposed registration-guided consistency architecture fosters task-specificity within the synthetic and registration modules by applying identical deformation fields before and after synthesis, while enforcing output consistency through an alignment loss. Moreover, the synthetic module is designed to possess the capability of disentangling anatomical structures and specific styles across various modalities. An anatomy consistency loss is introduced to further compel the synthetic module to preserve geometrical integrity within latent spaces. Experiments conducted on both an in-house abdominal CECT-CT dataset and a publicly available pelvic MR-CT dataset have demonstrated the superiority of the proposed method.

Boosting Medical Image Synthesis via Registration-guided Consistency and Disentanglement Learning

TL;DR

This work tackles misalignment noise in medical image synthesis by integrating a registration-guided consistency framework with an anatomy consistency disentanglement synthetic module. It introduces a registration module (deformation generator and resampler ) paired with an alignment loss to enforce geometry preservation across synthesis before and after applying the same deformation field, and an ACDS module that disentangles anatomical content from modality styles using dedicated encoders and generators, guided by adversarial, self-reconstruction, cycle-consistency, and anatomy-consistency losses. Experimental results on an in-house abdominal CECT-CT dataset and the SynthRAD2023 pelvic MR-CT dataset demonstrate superior quantitative performance (MAE, PSNR, SSIM) and clearer anatomical boundaries, with ablations confirming the contributions of both registration-guided consistency and ACDS. Overall, the approach provides a practical, geometry-preserving framework for robust cross-modality medical image synthesis, reducing misalignment artifacts and improving structural fidelity for clinical tasks such as MR-to-CT and CECT-to-CT synthesis.

Abstract

Medical image synthesis remains challenging due to misalignment noise during training. Existing methods have attempted to address this challenge by incorporating a registration-guided module. However, these methods tend to overlook the task-specific constraints on the synthetic and registration modules, which may cause the synthetic module to still generate spatially aligned images with misaligned target images during training, regardless of the registration module's function. Therefore, this paper proposes registration-guided consistency and incorporates disentanglement learning for medical image synthesis. The proposed registration-guided consistency architecture fosters task-specificity within the synthetic and registration modules by applying identical deformation fields before and after synthesis, while enforcing output consistency through an alignment loss. Moreover, the synthetic module is designed to possess the capability of disentangling anatomical structures and specific styles across various modalities. An anatomy consistency loss is introduced to further compel the synthetic module to preserve geometrical integrity within latent spaces. Experiments conducted on both an in-house abdominal CECT-CT dataset and a publicly available pelvic MR-CT dataset have demonstrated the superiority of the proposed method.
Paper Structure (19 sections, 12 equations, 3 figures, 2 tables)

This paper contains 19 sections, 12 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Registration-guided Consistency
  4. Registration-guided Module.
  5. Synthesis Before Registration.
  6. Synthesis After Registration.
  7. Alignment Loss.
  8. Anatomy Consistency Disentanglement Synthetic Module
  9. Synthetic process.
  10. Disentanglement Learning.
  11. Experiments
  12. Experimental Setup.
  13. Datasets.
  14. Implementation Details
  15. Evaluation Metrics.
  16. ...and 4 more sections

Figures (3)

  • Figure 1: The misaligned challenges in medical image synthesis. (a) Misaligned anatomical structures result in inaccurate mapping in supervised paradigms. (b) The previous registration-guided synthesis kong2021breaking still retains misalignment-induced noise and suffers from blurred boundaries due to the averaging effect.
  • Figure 2: Overview of our proposed methods. (a) The registration-guided consistency architecture. $\mathcal{L}_{align}$ is used to minimize the synthetic images from synthesis after registration and synthesis before registration. (b) The anatomy consistency disentanglement synthetic (ACDS) module. (c) Testing phase for medical image synthesis.
  • Figure 3: Visual comparison of synthesized results on CECT-CT and MR-CT datasets. The pink boxes show the structural details where misalignment are often encountered during training.