3D multi-modality Transformer-GAN for high-quality PET reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Positron emission tomography (PET) scans are crucial for diagnosing cellular metabolic abnormalities.
• Standard-dose PET (SPET) offers more diagnostic information than low-dose PET (LPET) but involves higher radiation risks.
• Reconstructing high-quality SPET images from LPET data is essential for reducing patient radiation exposure.

Purpose of the Study:

• To develop a novel method for reconstructing high-quality SPET images from LPET and T1-weighted MRI (T1-MRI) data.
• To reduce radiation dose associated with SPET imaging while preserving diagnostic accuracy.
• To leverage multi-modal data and advanced deep learning techniques for improved image reconstruction.

Main Methods:

• A 3D multi-modality edge-aware Transformer-GAN was proposed, integrating LPET and T1-MRI data.
• Separate CNN-based encoders extracted local features from each modality, followed by a multimodal feature integration module.
• A Transformer-based encoder captured global semantic information, and a CNN decoder generated SPET images, enhanced by an edge-aware loss function.

Main Results:

• The proposed method effectively reconstructed high-quality SPET images from LPET and T1-MRI.
• Experiments demonstrated superior performance compared to state-of-the-art methods in both qualitative and quantitative evaluations.
• The edge-aware loss function successfully retained critical edge detail information in the reconstructed images.

Conclusions:

• The developed Transformer-GAN effectively reconstructs high-quality SPET images, offering a promising solution for dose reduction in PET imaging.
• The multi-modal approach integrating LPET and T1-MRI enhances reconstruction accuracy and preserves diagnostic information.
• This method holds potential for improving clinical patient diagnosis by providing high-quality PET imaging with reduced radiation risks.