Iodinated contrast agents (ICAs) are essential in computed tomography angiography (CTA) for highlighting vascular structures. However, their clinical application is facing growing limitations due to the risks of contrast-induced nephropathy. Low-dose contrast agent CTA (LDCTA) vascular imaging provides a safer option by utilizing a smaller amount of contrast agent. Current LDCTA vascular imaging methods face challenges in maintaining stable vessel intensity and differentiating vessel boundaries from the surrounding anatomical background due to the complex topological structure. To address these challenges, we propose LPGNet, an unpaired vascular reconstruction framework that integrates Contrastive Language-Image Pretraining (CLIP) with anatomical intensity consistency constraints. The CLIP prior facilitates both differentiation between low-contrast/well-enhanced vascular images and identification of fine-grained structural details, thereby enhancing the optimization of the enhancement network. Learnable prompts guide CLIP to capture contrast-specific visual semantics, while an intensity constraint loss ensures enhanced aortic regions maintain clinically preferable intensity ranges. An iterative optimization strategy further refines both the enhancement network and prompts through progressive semantic alignment. Extensive experiments on both simulated and clinical datasets demonstrate that LPGNet outperforms state-of-the-art paired and unpaired methods in terms of vascular contrast, anatomical detail preservation, and diagnostic utility. These results highlight its potential as a practical solution for contrast-limited vascular imaging scenarios.
张跃怀 et al. (2026) studied this question.