Structure-aware vessel enhancement network for low-dose contrast agent CT angiography imaging

CT angiography (CTA) is essential for early diagnosis, preoperative assessment, and postoperative monitoring of vascular conditions. Traditional CTA depends on substantial amounts of contrast agents to obtain adequate vascular differentiation, potentially leading to contrast-induced nephropathy and adverse reactions. While low-dose contrast techniques reduce patient risk, they often degrade image quality, specifically impairing the detection of intricate, small vessels, thus restricting their clinical usefulness. To address this challenge, we propose a novel low-dose agent CTA (LDCTA) image enhancement network that integrates a structure-aware perceptual loss module with an adaptive deformable convolution module to improve vascular detail reconstruction under low-dose agent conditions. The perceptual loss utilizes a pre-trained vascular segmentation model to focus on anatomical areas, improving semantic coherence and structural accuracy. In addition, the deformable convolution module dynamically adjusts convolution kernel shapes based on local structures, improving feature extraction for irregular and small-scale vessels. The proposed method has been thoroughly validated on head-neck and thoracic datasets, with experimental results demonstrating superior image enhancement quality and vascular structure preservation compared to existing approaches.