Generation of¹⁸F-FDG PET standard scan images from short scans using cycle-consistent generative adversarial network

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Nuclear Medicine

Background:

• Positron emission tomography (PET) imaging quality is often limited by scan duration.
• Reducing scan time is crucial for patient comfort and throughput but can compromise image quality due to increased noise.
• Developing methods to reconstruct high-quality images from accelerated PET scans is an active area of research.

Purpose of the Study:

• To improve PET image quality by generating images comparable to standard scan durations from significantly shorter scans (1/8 and 1/16).
• To quantitatively and qualitatively assess the performance of a cycle-GAN model in reconstructing PET images.
• To explore the impact of patient Body Mass Index (BMI) on the cycle-GAN model's performance.

Main Methods:

• Utilized whole-body PET scans from 42 patients (41 18F-FDG, 1 68Ga-PSMA).
• Trained multiple cycle-GAN network instances on subsets of data stratified by mean patient BMI.
• Evaluated model performance using quantitative metrics (PSNR, SSIM, NRMSE, SUVmean, SUVmax) and qualitative assessment by nuclear medicine specialists.
• Tested model generalizability on a 68Ga-PSMA scan.

Main Results:

• The cycle-GAN model significantly improved quantitative metrics (PSNR, SSIM, NRMSE) for both 1/8 and 1/16 scan durations across radiotracers.
• Qualitative analysis indicated that 1/8 scan duration reconstructions were more visually appealing than 1/16, despite slightly lower quantitative scores.
• The cycle-GAN approach outperformed traditional NLM denoising in both image quality improvement and speed.
• Results were statistically significant (p < 0.05).

Conclusions:

• Cycle-GAN effectively reduces noise in accelerated PET scans, yielding quantitative and qualitative image quality comparable to standard scan durations.
• The model demonstrates consistent performance for both 18F-FDG and 68Ga-PSMA, though further study is needed for 68Ga-PSMA.
• Reconstructions from 1/8 scan duration inputs offer a favorable balance between image quality and scan time compared to 1/16 duration.