A groundbreaking study published in BME Frontiers has unveiled Ves-GAN, an unsupervised vessel-targeted denoising framework specifically developed to enhance low-dose coronary computed tomography angiography (LDCTA) imaging. This innovative approach tackles a pivotal challenge in cardiovascular diagnostics: achieving a balance between radiation safety and image clarity. By effectively minimizing noise while preserving intricate vascular structures, Ves-GAN represents a significant advancement in the field.
Coronary CTA is a cornerstone of noninvasive cardiovascular disease diagnosis. However, to limit radiation exposure, clinicians often employ LDCTA, which introduces noise and blurring that compromise diagnostic accuracy. Traditional denoising methods, such as iterative reconstruction, often fail to retain subtle vascular details. Supervised deep learning models, while effective, require paired noisy and clean datasets, which are clinically impractical. Ves-GAN circumvents this limitation through its unsupervised design, eliminating the need for paired data while maintaining structural integrity.
Ves-GAN builds on the Cycle-GAN architecture, integrating three key innovations. First, a high-frequency-aware data augmentation strategy applies localized filtering and varied noise patterns to LDCTA images, enhancing model robustness across diverse noisy conditions. Second, the high-frequency squeeze-and-excitation (HFSE) module incorporates Laplacian and Sobel operators to amplify sensitivity to fine vascular features. Third, a vessel-consistency loss function combines vessel-preserving and edge-preserving constraints, ensuring structural fidelity during denoising.
In a clinical study involving 50 CT scans, Ves-GAN outperformed existing unsupervised models. Quantitative assessments revealed average improvements of 7.5% in peak signal-to-noise ratio (PSNR) and 10.2% in structural similarity index (SSIM). Radiologists rated Ves-GAN highly in vascular clarity (4.8/5), edge visibility (4.9/5), and noise reduction (4.6/5), with results approaching the quality of normal-dose CTA images. These improvements directly enhance diagnostic reliability, particularly in detecting subtle lesions.
The development of Ves-GAN underscores the potential of AI-driven unsupervised methods in medical imaging. By prioritizing vascular structure preservation and leveraging high-frequency feature sensitivity, this framework sets a new standard for LDCTA denoising. As cardiovascular disease remains a leading global health concern, tools like Ves-GAN promise to enhance diagnostic precision, ultimately improving patient outcomes through more reliable and accessible imaging.
