Safm-Cta: Segmentation-Augmented Foundation Model for CT-to-CTA Translation

Computed Tomography Angiography (CTA) synthesis from non-contrast Computed Tomography (CT) scans offers a contrast-free alternative for vascular assessment, particularly benefiting patients at risk from iodinated contrast agents. However, existing generative models such as GANs or diffusion networks often fail to maintain vascular continuity and anatomical realism. In this study, we propose SAFM-CTA, a Segmentation-Augmented Foundation Model for high-fidelity CT-to-CTA translation. SAFM-CTA integrates a pretrained 3D Vision Transformer encoder that captures global anatomical priors with a multitask decoder jointly trained for CTA synthesis and vascular segmentation. Through segmentation-guided optimization and vessel-weighted content loss, the model enforces explicit anatomical consistency during generation. Evaluated on 1,886 paired CT-CTA scans, SAFM-CTA surpasses state-of-the-art methods (PSNR: 29.37, SSIM: 0.915), yielding superior vessel integrity. These results highlight the potential of SAFM-CTA as a clinically applicable foundation model enabling contrastfree angiographic imaging with strong structural coherence and perceptual fidelity.