Context-Aware Transformer GAN for Direct Generation of Attenuation and Scatter Corrected PET Data

Area of Science:

• Medical Imaging
• Deep Learning
• Positron Emission Tomography (PET)

Background:

• Positron Emission Tomography (PET) imaging requires attenuation and scatter correction (ASC) for accurate image reconstruction.
• Traditional ASC methods rely on transmission scans, which are not always available, especially with integrated PET/MRI systems.

Purpose of the Study:

• To develop and evaluate a deep learning framework for direct generation of ASC PET images from non-ASC (NASC) images.
• To compare the performance of different conditional generative adversarial network (cGAN) architectures, including Pix2Pix, AG-Pix2Pix, ViT-GAN, and Swin-GAN.

Main Methods:

• Trained cGAN models using single-modality (NASC) or multimodality (NASC+MRI) input data.
• Evaluated four cGAN models on retrospective ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) PET images from 33 subjects.
• Utilized quantitative metrics such as PSNR, MS-SSIM, NRMSE, and MAE for image quality assessment.

Main Results:

• No significant impact of input type (single vs. multimodal) on image quality metrics was observed.
• Swin-GAN showed superior performance in MS-SSIM and comparable or better results in PSNR, NRMSE, and MAE compared to other models when using multimodal data.
• The cGAN models, particularly Swin-GAN, consistently generated reliable and accurate ASC PET images.

Conclusions:

• The proposed context-aware generative deep learning framework effectively produces ASC PET images from NASC data.
• Swin-GAN is a highly effective model for this task, offering robust performance with both single- and multimodal inputs.
• This methodology enables ASC PET image generation without the need for transmission scans, beneficial for standalone PET or PET/MRI systems.