Syn2Real: synthesis of CT image ring artifacts for deep learning-based correction.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Ring artifacts degrade X-ray computed tomography (CT) image quality, limiting clinical utility.
• Current deep learning methods require large, high-quality datasets that are expensive and time-consuming to generate.
• Synthesizing training data in the image domain offers a scalable solution without relying on specific imaging system physics.

Purpose of the Study:

• To develop a novel, computationally efficient pipeline, 'Syn2Real,' for synthesizing realistic ring artifacts directly in the image domain.
• To enable scalable production of training data for deep learning-based ring artifact correction methods.
• To demonstrate the generalizability of synthetic data-trained models across different CT systems and imaging parameters.

Main Methods:

• Developed 'Syn2Real,' an image-domain pipeline for realistic ring artifact synthesis.
• Trained UNet, UNetpp (with self-attention), and diffusion models on synthetic and real CT data.
• Evaluated model performance on energy-integrating CT images and prototype photon-counting CT data with varying parameters.

Main Results:

• Models trained on synthetic data demonstrated effective ring artifact correction on diverse photon-counting CT images.
• Successful generalization across different energy levels and slice thicknesses, validating the synthetic data's realism.
• The approach proved versatile for various deep learning architectures and loss functions.

Conclusions:

• The Syn2Real pipeline provides a robust foundation for generating versatile training data for CT ring artifact correction.
• This method facilitates the development of more adaptable and effective artifact correction solutions for various CT applications.
• The study highlights the potential of image-domain data synthesis for advancing deep learning in medical imaging.