Curvilinear Structure-preserving Unpaired Cross-domain Medical Image Translation

TL;DR

Comprehensive evaluation across three imaging modalities: optical coherence tomography angiography, color fundus and X-ray coronary angiography demonstrates that CST improves translation fidelity and achieves state-of-the-art performance.

Abstract

Unpaired image-to-image translation has emerged as a crucial technique in medical imaging, enabling cross-modality synthesis, domain adaptation, and data augmentation without costly paired datasets. Yet, existing approaches often distort fine curvilinear structures, such as microvasculature, undermining both diagnostic reliability and quantitative analysis. This limitation is consequential in ophthalmic and vascular imaging, where subtle morphological changes carry significant clinical meaning. We propose Curvilinear Structure-preserving Translation (CST), a general framework that explicitly preserves fine curvilinear structures during unpaired translation by integrating structure consistency into the training. Specifically, CST augments baseline models with a curvilinear extraction module for topological supervision. It can be seamlessly incorporated into existing methods. We integrate it into CycleGAN and UNSB as two representative backbones. Comprehensive evaluation across three imaging modalities: optical coherence tomography angiography, color fundus and X-ray coronary angiography demonstrates that CST improves translation fidelity and achieves state-of-the-art performance. By reinforcing geometric integrity in learned mappings, CST establishes a principled pathway toward curvilinear structure-aware cross-domain translation in medical imaging.

Figures (9)

• Figure 1: Illustration of the vessel fidelity issue in cross-domain translation. The black or green indicating vessels from the source and translated images respectively. Previous methods show extra/missing or shifted vessels in translated results, represented by vessels in cyan. Our method preserves vessel structures well, making it more suitable for downstream segmentation tasks.
• Figure 2: Overview of our proposed method. The complete training workflow. The $A$ is the image from source domain and the $\tilde{A}$ denotes the image of $A$ translated from the source domain to the target domain. The baseline pipeline illustrating the baseline model’s original forward process.
• Figure 3: Architecture of the Curvilinear Extraction Module (CEM). We utilize the mask decoder output with sigmoid activation as CEM, which captures curvilinear structural semantics rather than low-level textural patterns from intermediate features.
• Figure 4: Qualitative results on the OCTA translation task (OCTA500$\leftrightarrow$ROSE). The OCTA500 to ROSE case corresponds to an eye with diabetic retinopathy, while the ROSE to OCTA500 case corresponds to a healthy eye. Other methods tend to distort curvilinear structures during translation, while ours preserves them more effectively.
• Figure 5: Qualitative results on the fundus translation task (DRIVE$\leftrightarrow$STARE). In particular, the regions highlighted by green boxes and arrowheads show that our model better maintains the original curvilinear structures from the source images.