Ves-GAN: Unsupervised Vessel-Targeted Low-Dose Coronary Computed Tomography Angiography Denoising Framework

Objective: This study aims to develop an unsupervised denoising framework for low-dose coronary computed tomography (CT) angiography (LDCTA), which reduces noise while preserving vascular structures. Impact Statement: This work proposes Ves-GAN, a novel denoising framework that meets the challenges of data acquisition and assumptions about noise characteristics. By providing robust noise reduction while maintaining vascular integrity, Ves-GAN facilitates more reliable clinical evaluations and improves the overall quality of cardiovascular diagnosis. Introduction: LDCTA minimizes radiation exposure in cardiovascular imaging but introduces noise and blurring, affecting diagnostic accuracy. Existing denoising methods, such as supervised deep learning models, require paired datasets and rely on noise assumptions. Unsupervised models show promise but often fail to preserve vascular structures, limiting clinical application. Methods: Ves-GAN incorporates a high-frequency-aware data augmentation strategy for robust generalization. The generator employs a high-frequency squeeze-and-excitation module to improve sensitivity to fine vascular features. Additionally, a vessel-consistency loss is introduced to preserve structural integrity during the denoising process. Results: On average, Ves-GAN achieves 7.5% and 10.2% improvements in peak signal-to-noise ratio and structural similarity index metrics compared to existing unsupervised models. Clinical validation involved 50 CT scans reviewed by 3 radiologists, who noted substantial enhancements in vascular clarity and lesion visibility. Conclusion: Ves-GAN outperforms existing unsupervised models in preserving vascular details and noise reduction, significantly enhancing clinical diagnostic reliability.