High-quality PET image synthesis from ultra-low-dose PET/MRI using bi-task deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Positron emission tomography (PET) imaging quality is often compromised at lower radiation doses, leading to increased noise and reduced detail.
• Reducing radiation exposure is crucial for patient safety in medical imaging procedures.
• Current methods struggle to balance image quality with minimal radiation burden.

Purpose of the Study:

• To develop a method for acquiring high-quality PET images from ultra-low-dose states.
• To achieve both high-quality imaging and a reduced radiation burden for patients.
• To introduce a novel deep learning model for enhanced PET image reconstruction.

Main Methods:

• A two-task-based end-to-end generative adversarial network (bi-c-GAN) was developed, integrating PET and MRI data.
• A combined loss function (mean absolute error, structural loss, bias loss) was employed to optimize the model.
• The model was trained and validated using integrated PET/MRI data from 67 patients.

Main Results:

• The bi-c-GAN model outperformed U-net, c-GAN, and multiple input c-GAN in quantitative metrics like PSNR, SSIM, NMSE, and CNR.
• Bi-c-GAN achieved at least 6.7% higher PSNR and 8% higher CNR in 5% low-dose PET imaging compared to other methods.
• The model demonstrated significant denoising and image quality improvement in ultra-low-dose PET scans.

Conclusions:

• The proposed bi-c-GAN effectively enhances ultra-low-dose PET image quality using integrated PET/MR data and multitask deep learning.
• This method offers a promising approach to reduce radiation exposure while maintaining diagnostic image quality.
• Bi-c-GAN represents an advancement in medical imaging AI for safer and more effective PET scans.