Low-count whole-body PET denoising with deep learning in a multicenter, multi-tracer and externally validated study

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Positron Emission Tomography (PET) is crucial but limited by cost, accessibility, and radiation exposure.
• Deep learning (DL) denoising offers a solution for low-count PET scans, potentially reducing tracer dose or scan time.
• Clinical validation of DL denoising across diverse scanners and radiotracers is essential.

Purpose of the Study:

• To evaluate the clinical performance of a DL-based denoising software (NUCLARITY) in a multicenter setting.
• To assess image quality, diagnostic confidence, and lesion detection using denoised low-count PET scans.
• To validate the generalizability of the DL algorithm across different PET scanners and radiotracers.

Main Methods:

• A multicenter, blinded evaluation of NUCLARITY using 65 PET scans ([18F]FDG, [18F]PSMA, [68Ga]PSMA, [68Ga]DOTATATE) from GE and Siemens systems.
• Simulated 50% low-count scans were denoised and compared to standard-count scans using quantitative metrics (RMSE, PSNR, SSIM) and reader assessments.
• Nuclear physicians evaluated diagnostic image quality, confidence, lesion detection, and quantification (SUVmean, SUVmax, MTV).

Main Results:

• Denoised low-count enhanced (LCE) scans demonstrated superior quantitative image quality compared to unenhanced low-count scans.
• High concordance (CCC=1.00 for SUVmean, 0.99 for SUVmax) was observed between standard-count (SC) and LCE scans for lesion quantification.
• LCE scans showed 99% sensitivity and 99% specificity for lesion detection, with diagnostic performance comparable to SC scans, despite slight decreases in perceived quality and confidence.

Conclusions:

• This study provides the first blinded, multicenter reader evaluation of a PET denoising algorithm in Europe across multiple tracers and unseen scanner technologies.
• The NUCLARITY algorithm showed robust generalizability and preserved diagnostic accuracy with 50% reduced PET scan counts.
• Findings support the clinical integration of DL-based PET denoising to enable reduced radiation dose or scan time for common radiotracers.