Sinogram Inpainting with Generative Adversarial Networks and Shape Priors.

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• X-ray computed tomography (CT) reconstructs images from X-ray absorption profiles (sinograms).
• Image reconstruction is an ill-posed inverse problem, especially with insufficient X-ray measurements, leading to artifacts.
• Limited angle CT scans, where data is missing from certain directions, pose a significant challenge.

Purpose of the Study:

• To develop a novel method for reducing image artifacts in limited-angle X-ray CT.
• To infer missing sinogram data using prior information about the object's shape.
• To improve image quality in scenarios with substantial, consecutive missing tomographic measurements.

Main Methods:

• A Generative Adversarial Network (GAN) was employed to combine limited acquisition data with shape priors.
• The method focuses on inferring consecutive missing X-ray measurements, unlike previous techniques.
• The approach was evaluated against state-of-the-art sinogram inpainting methods.

Main Results:

• The proposed method consistently improved image quality compared to existing techniques.
• A significant 7 dB Peak Signal-to-Noise Ratio (PSNR) improvement was demonstrated.
• The GAN-based approach effectively reduced image artifacts caused by limited tomographic data.

Conclusions:

• Shape-prior-guided sinogram inpainting using GANs is effective for limited-angle CT.
• The method offers a robust solution for reconstructing high-quality CT images from incomplete datasets.
• This approach advances the field of CT image reconstruction, particularly for challenging acquisition geometries.