Deep learning-based low count whole-body positron emission tomography denoising incorporating computed tomography...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiochemistry

Background:

• Deep learning significantly improves image quality and quantification in low-count (LC) positron emission tomography (PET).
• Current deep learning methods for LC PET denoising typically use only a single LC PET image as input.
• There is a need to further reduce radiation dose levels in PET imaging.

Purpose of the Study:

• To develop and evaluate a deep learning-based denoising method for LC PET that incorporates computed tomography (CT) priors.
• To assess the effectiveness of this CT-prior-enhanced method in further reducing radiation dose levels while maintaining image quality and diagnostic accuracy.

Main Methods:

• Fifty patients undergoing 2-deoxy-2-[18F]fluoro-D-glucose (18F-FDG) PET/CT scans were retrospectively analyzed.
• Low-count (LC) PET data (LC-10 and LC-20) were generated by down-sampling full-count (FC) PET data.
• U-Net and cGAN models were trained with and without CT image incorporation for denoising.

Main Results:

• The deep learning models incorporating CT priors (U-Net-2, cGAN-2) outperformed those without (U-Net-1, cGAN-1) in denoising performance.
• cGAN-2 achieved the best results in terms of mean square error and structural similarity index.
• Models with CT priors (cGAN-2, U-Net-2) demonstrated lower errors in SUV quantification and improved lesion detectability, successfully retrieving lesions missed by conventional methods.

Conclusions:

• Deep learning-based LC PET denoising incorporating CT priors is more effective than conventional methods using single LC PET input.
• This approach significantly enhances image quality and lesion detectability, particularly at lower radiation dose levels.
• The integration of CT priors offers a promising strategy for dose reduction in PET imaging without compromising diagnostic performance.