Automatic Contrast Generation from Contrastless Computed Tomography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Contrast-enhanced computed tomography (CTCA) is standard for coronary artery disease (CAD) detection but involves contrast agents, radiation, and higher costs.
• Deep learning generative models offer a potential solution by creating pseudo-enhanced images from non-contrast CT scans.

Purpose of the Study:

• To evaluate generative adversarial networks (GANs), specifically Pix2Pix-GAN and Cycle-GAN, for generating pseudo-enhanced CTCA images from non-contrast CT.
• To explore the impact of 2D versus 3D data and architectures on GAN performance.

Main Methods:

• Paired non-contrast CT and CTCA scans from private and public datasets were used.
• Pix2Pix-GAN and Cycle-GAN models were trained and evaluated using metrics like SSIM, PSNR, and FID.
• An analysis of 2D vs. 3D input and architectural variations was conducted.

Main Results:

• Pix2Pix-GAN with 2D data achieved higher SSIM (0.492) and PSNR (16.375 dB) but produced blurred images.
• Cycle-GAN models generated visually clearer images compared to Pix2Pix-GAN.
• Fréchet Inception Distance (FID) was crucial for evaluating image quality beyond traditional metrics, highlighting Cycle-GAN's advantage in visual fidelity.

Conclusions:

• Generative models show promise for reducing CTCA risks and costs, particularly for screening asymptomatic individuals or ruling out CAD in acute settings.
• Cycle-GAN appears more suitable for clinical translation due to superior visual quality, despite potentially lower scores on SSIM/PSNR alone.
• Further validation is needed, but this approach could significantly impact CAD detection strategies.