Medical Radiation Exposure Reduction in PET via Super-Resolution Deep Learning Model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiochemistry

Background:

• Positron emission tomography (PET) image quality is dependent on injected dose and scan duration.
• Reducing scan time or radiotracer dose can compromise diagnostic image quality.
• Super-resolution (SR) deep learning offers a potential method to improve image quality from reduced data acquisition.

Purpose of the Study:

• To evaluate if SR deep learning can enhance 50% acquisition time PET images to the quality of 100% acquisition time images.
• To determine the feasibility of reducing injected [18F]-fluorodeoxyglucose (FDG) dose using SR deep learning.
• To assess the impact of SR deep learning on image quality metrics and subjective observer scores.

Main Methods:

• Retrospective analysis of PET data from 108 adult patients.
• Application of a supervised deep learning SR technique to PET images acquired with 50% of the standard time.
• Nested cross-validation using nine data subsets.
• Objective image quality assessment using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM).
• Subjective image quality assessment using mean opinion scores (MOS) by observers.

Main Results:

• SR-processed 50% acquisition time PET images (SR-PET) showed improved objective quality (PSNR: 31.3 dB, SSIM: 0.931) compared to standard 50% PET.
• Subjective MOS scores for SR-PET (lung: 3.96, liver: 3.80, bowel: 3.67) were significantly higher than for 50% PET (lung: 3.41, liver: 3.31, bowel: 3.08).
• SR-PET image quality approached that of 100% acquisition time PET images (lung: 4.23, liver: 4.27, bowel: 3.94).

Conclusions:

• SR deep learning effectively improves the image quality of 50% acquisition time PET scans.
• SR-PET images are subjectively and objectively superior to standard 50% PET images.
• This technique holds promise for reducing injected FDG doses and associated radiation exposure in PET imaging.