Generative Adversarial Network of Industrial Positron Images on Memory Module

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Positron Emission Computed Tomography (PET) imaging faces challenges due to its ill-posed nature and limited photo response data.
• Generative adversarial networks (GANs) have shown significant success in computer vision tasks.

Purpose of the Study:

• To improve the quality of industrial PET images using an attention-based generative adversarial network.
• To introduce a novel memory module for focusing on image feature details.

Main Methods:

• Training an adversarial model incorporating an attention mechanism for PET image enhancement.
• Utilizing an encoder to extract hidden vectors from a basic dataset as prior knowledge.
• Jointly training the generative and discriminative networks.

Main Results:

• The proposed method significantly improved the quality of simulated PET images, as validated by MS-SSIM and PSNR metrics.
• Real industrial PET images processed with the model demonstrated a good visual effect.

Conclusions:

• The attention-based GAN with a memory module is effective for enhancing industrial PET image quality.
• This approach offers a promising solution for overcoming limitations in PET imaging data and resolution.