VesselTransGAN to CT Imaging: A Contrast Medium Free CTA Solution

Abstract

Computed Tomography Angiography (CTA) is vital for detecting and planning treatment for vascular disease, but its use of contrast medium limits viability for patients with renal insufficiency. Based on the standard Computed Tomography (CT) scans, which did not require injected contrast medium, the cross-modal conversion technology generates CT-CTA images that provide reference properties equivalent to CTA diagnosis. Our research presented VesselTransGAN, a framework designed to create CTA from CT scans by leveraging the Generative Adversarial Network (GAN) framework and 2D-3D fusion strategy. This innovative approach comprised two key elements. Firstly, it employed pixel grayscale alignment to enhance blood vessels locally. Secondly, a 2D-3D fusion strategy improved each slice’s image quality and continuity among slices. Through quantitative evaluation of image quality, compared with SOTA models, our model has improved performance in MSE, PSNR, and SSIM. We also found that nearly 90% of the generated CTAs were of medium or high quality as assessed by clinicians. Upon comparing the synthesized CTA and authentic CTA assessments conducted by clinicians, our framework showed promising potential in diagnosing test sets on 426 paired CT-CTA images from three major medical centers, demonstrating improved results compared to the SOTA model.VesselTransGAN has the potential to improve the safety and accessibility of CTA scans for patients with renal insufficiency and to enhance the accuracy of CTA scans for diagnosing vascular disease. We release our code at https://github.com/Flora-huay/VesselTranGAN.

Keywords

• CTA
• clinical diagnosis
• generative model
• image generation