Generative adversarial network-based sinogram super-resolution for computed tomography imaging

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Computed tomography (CT) 2x2 acquisition mode offers improved projection efficiency and reduced X-ray exposure compared to 1x1 mode.
• However, the 2x2 mode yields low-resolution (LR) sinograms, resulting in poor reconstructed image quality, limiting its clinical application.
• Super-resolution (SR) techniques are crucial for enhancing LR data in medical imaging.

Purpose of the Study:

• To propose a novel sinogram-super-resolution (SR) generative adversarial network (GAN) model for improving CT image reconstruction quality using the 2x2 acquisition mode.
• To generate high-resolution (HR) sinograms from LR sinograms acquired via the 2x2 mode.
• To enable the wider adoption of the efficient 2x2 acquisition mode in CT systems.

Main Methods:

• A novel SR generative adversarial network (GAN) model was developed, featuring a residual network generator for LR sinogram feature extraction and SR sinogram generation.
• A relativistic discriminator was employed to enhance the realism of the generated SR sinograms.
• The model's training incorporated a combined loss function including cycle consistency loss, sinogram domain loss, and reconstruction image domain loss.

Main Results:

• The proposed GAN model successfully generated clean, high-resolution (HR) sinograms from noisy low-resolution (LR) sinograms.
• Qualitative and quantitative evaluations on both digital and real CT data demonstrated the model's effectiveness.
• The generated SR sinograms significantly improved the quality of CT images reconstructed using the filtered-back-projection algorithm.

Conclusions:

• The developed sinogram-super-resolution GAN model effectively addresses the resolution limitations of CT's 2x2 acquisition mode.
• This approach enhances CT image quality while maintaining the benefits of reduced X-ray exposure and improved data collection efficiency.
• The proposed method shows significant potential for advancing CT imaging systems by enabling the use of more efficient acquisition modes.