An Innovative Low-dose CT Inpainting Algorithm based on Limited-angle Imaging Inpainting Model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Increasing popularity of computed tomography (CT) scans necessitates strategies for radiation dose reduction.
• Public concern regarding CT radiation exposure is rising.
• Balancing diagnostic image quality with reduced radiation dose is a key research focus.

Purpose of the Study:

• To demonstrate the efficacy of limited-angle CT imaging for acquiring lower-dose CT images.
• To introduce a generative adversarial network-based image inpainting model, LDLAIM, for restoring low-dose CT images with limited-angle acquisition.

Main Methods:

• Development of the Low-dose imaging and Limited-angle imaging inpainting Model (LDLAIM) using generative adversarial networks (GANs).
• Utilized three datasets (chest and abdomen, head, phantom) synthesized with low-dose and limited-angle CT parameters for training.
• Employed a generator with residual blocks, an encoder-decoder structure, and skip connections, alongside a discriminator for network training.

Main Results:

• The LDLAIM model achieved high performance across three datasets, with mean SSIM values of 0.984, 0.981, and 0.977.
• Quantitative metrics (SSIM, PSNR, RMSE) demonstrated the superiority of LDLAIM over other algorithms.
• Subjective quality scores also indicated the proposed method's effectiveness in restoring CT images.

Conclusions:

• The proposed LDLAIM method effectively restores CT images acquired using simultaneous low-dose and limited-angle techniques.
• Limited-angle CT imaging is a viable technique for reducing CT radiation dose.
• This research offers a novel approach for low-dose CT imaging studies.