Deep learning-assisted ultra-fast/low-dose whole-body PET/CT imaging

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Artificial Intelligence

Background:

• Reducing radiation hazards and improving patient comfort in PET scans are key goals.
• Current methods involve moderating injected activity or reducing acquisition time.
• Deep learning offers potential for synthesizing high-quality PET images from reduced-dose scans.

Purpose of the Study:

• To assess the performance of deep learning models in synthesizing regular full-dose (FD) whole-body (WB) PET images from fast/low-dose (LD) acquisitions.
• To evaluate the diagnostic quality and quantitative accuracy of synthesized FD PET images.

Main Methods:

• Two separate WB 18F-FDG PET/CT studies (FD ~27 min, LD ~3 min) were acquired from 100 patients.
• Modified CycleGAN (CGAN) and ResNET (RNET) models were used to predict FD PET images from LD scans.
• Image quality was assessed by nuclear medicine physicians, and lesion detectability was evaluated using a pass/fail scheme. Quantitative analysis included SUV bias for various organs and malignant lesions.

Main Results:

• CGAN achieved adequate to good qualitative scores for brain (4.92/5) and neck+trunk (3.88/5).
• Average SUV bias over normal tissues was 3.39 ± 0.71% for CGAN and -3.83 ± 1.25% for RNET.
• CGAN demonstrated the lowest SUV bias (0.01%) with a 95% CI of -0.36, +0.47 for malignant lesions compared to reference FD images.

Conclusions:

• CycleGAN can effectively synthesize clinical FD WB PET images from LD acquisitions (1/8th standard time/activity).
• The synthesized FD images show comparable performance to reference FD images in lesion detectability, qualitative assessment, and quantitative accuracy.
• Deep learning-based synthesis offers a promising approach to reduce radiation dose and improve patient experience in PET imaging.