Promising Generative Adversarial Network Based Sinogram Inpainting Method for Ultra-Limited-Angle Computed Tomography...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Limited-angle computed tomography (CT) poses significant reconstruction challenges, especially in ultra-limited-angle scenarios (<90°).
• Traditional iterative reconstruction algorithms struggle to effectively address severe artifacts in ultra-limited-angle CT.
• Generative Adversarial Networks (GANs) show promise in image inpainting by restoring missing image information.

Purpose of the Study:

• To propose a novel deep learning approach for ultra-limited-angle CT image reconstruction.
• To develop a GAN-based method for restoring missing sinogram data and suppressing artifacts.
• To improve the quality of CT images reconstructed from severely limited projection data.

Main Methods:

• A sinogram-inpainting-GAN (SI-GAN) was developed, incorporating a U-Net generator and a patch-design discriminator.
• A joint projection and image domain loss function, including back-projection weighted image domain loss, was utilized.
• The network was trained using paired limited-angle and 180° sinograms to learn sinogram data continuity.

Main Results:

• The proposed SI-GAN method successfully restored missing sinogram data for ultra-limited-angle CT.
• The technique effectively suppressed artifacts commonly associated with ultra-limited-angle scanning.
• Both simulation studies and experiments with actual data demonstrated the method's efficacy.

Conclusions:

• The SI-GAN method offers a robust solution for ultra-limited-angle CT image reconstruction.
• This deep learning approach significantly mitigates artifacts, enhancing image quality in challenging CT scenarios.
• The study highlights the potential of GANs for addressing data-deficiency problems in medical imaging reconstruction.