Deep learning-based partial volume correction in standard and low-dose positron emission tomography-computed...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Positron emission tomography (PET) imaging is limited by partial volume effects (PVEs) due to system resolution.
• PVEs cause significant bias and image blurring, especially in small structures.
• Current methods struggle to simultaneously correct PVEs and denoise low-dose PET images.

Purpose of the Study:

• To develop a deep learning framework for joint partial volume correction (PVC) and denoising of PET images.
• To predict partial volume corrected full-dose (FD + PVC) images from standard or low-dose (LD) PET data.
• To provide a solution without requiring anatomical information.

Main Methods:

• A modified encoder-decoder U-Net network was trained using LD or standard PET images as input.
• The network's target was FD + PVC images generated by six different PVC methods (GTM, MTC, RBV, IY, RVC, RL).
• Model performance was evaluated using PSNR, RMSE, SSIM, and bias metrics.

Main Results:

• The reblurred Van-Cittert (RVC) method achieved high structural similarity (SSIM) with LD (0.89) and FD (0.94) PET images.
• Iterative Yang (IY) and geometric transfer matrix (GTM) also showed promising results.
• Multi-target correction (MTC) and Richardson-Lucy (RL) methods exhibited larger quantitative errors.

Conclusions:

• The proposed deep learning framework effectively performs joint denoising and PVC for PET images.
• The models can enhance image quality for LD or standard PET-CT scans when anatomical data is unavailable.
• This approach offers a valuable tool for improving PET image analysis.