Noise and Signal Characteristics of Deep Learning-Based Denoising for a SiPM-based PET/CT Scanner

Objectives: Deep convolutional neural networks (DCNN) can be trained to adapt to different noise levels in input PET images and produce consistent denoised results across different patient studies. However, DCNN denoising is highly nonlinear and its signal and noise characteristics on real scans have not been fully investigated. The objectives of this study were to: 1. Evaluate the noise adaptive DCNN on a SiPM-based PET/CT scanner using both phantom and clinical studies. 2. Investigate its performance in terms of ensemble noise (reproducibility) and bias on real patients with different scan durations.

Methods: An eight-layer deep residual denoising network trained with 8 patient studies that was previously published was used in this study. We first evaluated its quantitative performance on a NEMA IQ phantom acquired on a SiPM-based PET/CT scanner that comes with less than 270ps time-of-flight resolution and 27cm axial FOV. The spheres and background concentrations were 16.8 kBq/mL and 4.33 kBq/mL, respectively, to achieve 4:1 contrast. The scatter phantom had 120 MBq activity. The scan was acquired for 243 seconds for each of the three bed positions so as to cover 100cm in 30 minutes with 50% overlap. The list-mode data was time-cut to generate 2-min and 1-min/bed datasets. We compared the OSEM reconstructed images with TOF and PSF modeling, the Gaussian post-filtered (GF) images (4mm FWHM) and the DCNN denoised images. Sphere contrast-recovery-coefficients (CRC) and background variability (BV) were used as figures of merit. We also acquired a patient study with 10-min/bed for three bed positions (~542 million total prompt coincidences), to approximately serve as a ground truth dataset. We evenly time-cut the list-mode data into 45s, 60s, 90s and 120s/bed datasets. For each of the scan durations, we generated 10 noise realizations using bootstrap. All datasets were reconstructed using OSEM, followed by GF (6mm FWHM) and DCNN. We computed ensemble bias using the reconstruction of the 10-min dataset as the approximate true mean in the 2 ROIs with elevated uptakes. We also computed both image roughness and ensemble noise for a background ROI in the liver.

Results: For the NEMA IQ phantom, DCNN yielded similar CRC as the input OSEM while reducing BV in the 4-min and 2-min studies. In the 1-min scan, DCNN produced similar BV to the 4-min and 2-min scans, however, it reduced the CRC on the 10-mm sphere due to smoothing. Compared to the GF results, DCNN of the 2-min scan yielded better CRC and BV than the 4-min GF results, which suggests that DCNN can potentially provide a 50% scan duration reduction. For Patient studies, DCNN of the 60s image is visually similar to the 600s OSEM. The bias maps show that DCNN yielded less bias than GF for the 90s and 120s scans, similar bias for the 60s scan, and higher bias in the lower uptake regions in the 45s scan. However, DCNN yielded consistent ensemble noise maps across all the scan durations. The ROI quantifications show that DCNN yielded similar bias as the input OSEM across all the scan durations except in the 45s image. It produced consistent image roughness and ensemble noise in the liver ROI across all the scans. The patient study also suggests that DCNN with 60s produced less bias in focal hot spots and noise than GF with 120s.

Conclusions: Ensemble bias and noise results show that DCNN can adapt to different noise levels in input images and produce consistent image quality across a wide range of count levels. Furthermore, DCNN can yield superior image quality in terms of ROI bias and background noise compared to the GF image under standard clinical protocols even when the scan duration is reduced by 50%. However, further reducing the scan duration may lead to higher bias in some regions even when the denoised image visually appears to be similar to the high-count study.