Contrast agent dose reduction in computed tomography with deep learning using a conditional generative adversarial...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Intravenous iodine-based contrast media (ICM) are crucial for CT imaging.
• High doses of ICM can pose risks and increase costs.
• Developing methods to reduce ICM dosage is an ongoing area of research.

Purpose of the Study:

• To investigate the efficacy of generative adversarial networks (GANs) in creating virtual contrast-enhanced CT images.
• To determine if GANs can reduce the required dose of ICM while maintaining diagnostic image quality and accuracy.

Main Methods:

• Dual-energy CT data from 85 patients were used to train and test four GAN models.
• Models were trained with reduced ICM levels (-50% and -80%) using 2D and 2.5D input formats.
• Image quality was assessed using L1 loss, SSIM, PSNR, and FID scores.
• Radiologists performed a visual Turing test to evaluate pathological consistency.

Main Results:

• GAN models achieved high image quality scores (SSIM >98%, PSNR >48) even with -80% ICM reduction.
• The -50% ICM reduction models demonstrated superior performance with SSIM >99% and PSNR >51.
• Only the -50% ICM reduction models achieved 100% pathological consistency, essential for clinical application.

Conclusions:

• Generative adversarial networks enable a 50% reduction in ICM for CT scans without compromising image quality or diagnostic accuracy.
• Pathological consistency is a critical factor for safety, and 80% ICM reduction may impact it.
• Further validation through phantom and animal studies is recommended.