Plasticine: A Traceable Diffusion Model for Medical Image Translation

Tianyang Zhanng, Xinxing Cheng, Jun Cheng, Shaoming Zheng, He Zhao, Huazhu Fu, Alejandro F. Frangi, Jiang Liu, Jinming DuanarXiv 2025 · DOI · arXiv

Introduction: The Traceability Gap in Medical Imaging

In medical diagnostics, translating images between modalities (e.g., MRI to CT) or predicting disease progression is vital. While Generative Adversarial Networks (GANs) and Diffusion Models produce high-fidelity synthetic images, they often operate as "black boxes." For clinicians, these models suffer from a critical flaw: a lack of traceability.

Standard diffusion models often lack spatial traceability, making it difficult to map specific anatomical features between modalities.

Most existing models change pixel intensities without accounting for anatomical movement. If a lesion appears in a translated image, a doctor must be able to verify its exact correspondence in the original scan. Without this spatial correspondence, synthetic images remain clinically untrustworthy.

The Plasticine Framework

To address this, the authors introduce Plasticine, a diffusion-based framework designed to decouple intensity shifts from spatial warping. Plasticine integrates two primary components within a single denoising routine:

  1. Diffusion Intensity Translation: Unlike standard models starting from Gaussian noise, Plasticine begins denoising from a noised version of the source image, guided by unsupervised "structure maps" to maintain basic content shapes.
  2. Cross-Modality Spatial Transformation: A dedicated module predicts diffeomorphic deformations ($\phi$). These are smooth, invertible, and topology-preserving mappings that ensure anatomical "stretching" remains biologically plausible.
The Plasticine framework allows for explicit tracking of anatomical parts (e.g., the heart) through learned deformation fields.

By learning the deformation field $\phi$ alongside intensity changes, Plasticine enables "interactive tracking." A clinician can select a region of interest in the source image and trace its exact transformation into the target modality, as shown in the heart translation example above.

Experimental Evaluation and Results

Plasticine was evaluated on three distinct tasks: Retinal OCT protocol translation, Chest MRI-to-CT, and Cardiac MRI pathology prediction (using datasets like UKBiobank and Learn2Reg).

Performance Metrics

The model consistently outperformed state-of-the-art baselines like CycleGAN, MUNIT, and DDIB:

  • Structural Accuracy: On Retinal OCT data, Plasticine achieved a Dice score of 0.801, far exceeding CycleGAN (0.624), indicating superior anatomical alignment.
  • Synthesis Quality: Plasticine recorded a Fréchet Inception Distance (FID) of 96.28, significantly better than MUNIT’s 119.77.

Clinical Validation

A study involving three clinicians confirmed that Plasticine’s generated lesions were more anatomically "reasonable" and provided significantly higher satisfaction regarding traceability compared to non-traceable baselines.

Conclusion and Future Work

Plasticine represents a significant step toward interpretable AI in medicine. By providing a mathematical bridge (the deformation field) between source and target images, it moves beyond simple image generation toward true anatomical translation. While currently limited to 2D images and requiring structure map pre-processing, future work aims to extend the framework to 3D volumes to better support full clinical workflows.

Summary updated

Abstract

Domain gaps arising from variations in imaging devices and population distributions pose significant challenges for machine learning in medical image analysis. Existing image-to-image translation methods primarily aim to learn mappings between domains, often generating diverse synthetic data with variations in anatomical scale and shape, but they usually overlook spatial correspondence during the translation process. For clinical applications, traceability, defined as the ability to provide pixel-level correspondences between original and translated images, is equally important. This property enhances clinical interpretability but has been largely overlooked in previous approaches. To address this gap, we propose Plasticine, which is, to the best of our knowledge, the first end-to-end image-to-image translation framework explicitly designed with traceability as a core objective. Our method combines intensity translation and spatial transformation within a denoising diffusion framework. This design enables the generation of synthetic images with interpretable intensity transitions and spatially coherent deformations, supporting pixel-wise traceability throughout the translation process.
BibTeX
@misc{zhanng2025plasticine,
  author = {Tianyang Zhanng and Xinxing Cheng and Jun Cheng and Shaoming Zheng and He Zhao and Huazhu Fu and Alejandro F. Frangi and Jiang Liu and Jinming Duan},
  title = {{Plasticine: A Traceable Diffusion Model for Medical Image Translation}},
  year = {2025},
  eprint = {2512.18455},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.18455},
}
Figure