Partial-ring PET image restoration using a deep learning based method.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Nuclear Medicine

Background:

• Partial-ring Positron Emission Tomography (PET) scanners offer potential benefits but introduce image artifacts due to incomplete projection data.
• Reconstructing high-quality images from partial-ring PET data remains a significant challenge in nuclear medicine.

Purpose of the Study:

• To investigate the efficacy of a deep learning (DL) based method for recovering images from partial-ring PET data.
• To compare the performance of DL in the projection domain versus the image domain for artifact reduction.

Main Methods:

• Simulated partial-ring PET data using the SimSET toolkit with varying numbers of removed detector blocks.
• Trained a convolutional neural network (CNN) based on a residual U-Net architecture to predict full-ring data from partial-ring data.
• Evaluated performance using Mean Squared Error (MSE), Structural Similarity Index (SSIM), and Recovery Coefficient (RC).

Main Results:

• The image-domain DL approach demonstrated superior artifact reduction (91.7% MSE reduction) compared to the projection-domain approach (14.3%).
• High image quality was maintained with SSIM values of 0.998, 0.996, and 0.993 for 3, 5, and 7 detector block removals, respectively.
• Accurate recovery of gray and white matter activity was achieved, with lesion recovery coefficients of 94%, 89%, and 79% for increasing detector block removals.

Conclusions:

• Deep learning, particularly the image-domain approach, shows significant potential for recovering high-quality PET images from partial-ring scanner data.
• This DL method can effectively mitigate artifacts and accurately quantify activity, even with considerable detector block loss.