3.2.1. Reconstruction-Based Noise Reduction Approaches
Ordered subset expectation maximization (OSEM) is a reconstruction algorithm that is generally used to reconstruct PET images [
42]. It is an iterative expectation maximization (EM) algorithm that tries to find the maximum likelihood (ML) solution, which corresponds to the most likely image given the measured projection data. During each iteration, a subset of the measured data is employed to estimate a new image with a higher likelihood. However, with each update of OSEM, the noise in the reconstructed image increases because of the noisy projection data. Therefore, one approach is to limit the number of updates during reconstruction limiting the reconstructed image noise. However, this means that full convergence is not guaranteed, which may result in an underestimation of radioactivity concentrations [
43,
44]. Another approach is to ensure full convergence of the reconstructed image by applying enough updates during the reconstruction, while a low-pass filter is applied to the final reconstructed image to reduce the impact of noise.
An approach that allows full convergence is the Bayesian penalized likelihood (BPL) or regularized iterative reconstruction algorithm. In comparison to a regular OSEM algorithm, a noise regularization or penalty term is added to the objective function with a weighting or beta value (β value). Higher β values depict a higher weighting of the regularization term resulting in a more effective noise suppression during the reconstruction. However, including a penalty term in the objective function generally needs an adaption of the optimization scheme as well, with Block-Sequential Regularized Expectation Maximization (BSREM), an optimizer that is frequently used for BPL reconstruction algorithms [
45,
46]. Based on the studied literature two BPL reconstruction algorithms have been investigated with respect to
68Ga-labeled radiotracers.
One of those BPL reconstruction algorithms was introduced by GE Healthcare under the name Q.Clear, which includes a noise regularization term based on relative differences between neighboring voxels and uses BSREM as an optimizer. In addition, Q.Clear incorporates both time-of-flight (TOF) information and point spread function (PSF) modeling. A literature search including both BSREM and Q.Clear revealed multiple studies on the optimal use of Q.Clear for different
68Ga-labeled radiotracers using both patient and phantom studies. These studies demonstrated that increasing the β value resulted in a reduction in the image noise [
26,
27,
28,
29,
30,
31,
32,
33], an improvement in the signal-to-noise ratio (SNR) [
26,
27,
29,
30,
32], and a reduction in Gibbs artifacts [
33] at the cost of a reduction in the maximum standardized uptake value (SUV
max) [
26,
28,
29,
31], CR [
26,
28,
30], RC [
33], and signal-to-background ratio (SBR) [
27,
29,
32]. Additionally, the mean standardized uptake value (SUV
mean) was almost identical for different β values [
26,
27,
28,
29]. Increasing the number of iterations had a minor effect on image noise in BSREM [
26]. It was further observed that the same β values have different impacts on SUV
max, CR, and SNR of lesions with different sizes and uptakes. Specifically, an increase in the β value enhanced the relative difference in SUV
max, CR, and SNR as the size and uptake of lesions decreased [
26,
29,
32].
Quantitative image analysis and visual assessment by experts confirmed that BSREM can outperform OSEM with respect to image quality. However, these findings were only valid for a certain range of β values. Different studies investigated the potential of BSREM in comparison to OSEM for different
68Ga-labeled radiotracers, while using different imaging settings and measures of image quality. As a result, different ranges of recommended β values were reported, as presented in
Table 3. BMI can also affect the optimal β value as Zanoni et al. suggested a rather large β value of 1600 for
68Ga-DOTANOC PET scanning of scan patients with 25 ≤ BMI < 30 or BMI ≥ 30 [
34]. However, these ranges of β values need to be interpreted with care since this is a GE proprietary and unitless tuning parameter, which can also depend on the version of the reconstruction software. Nevertheless, the different ranges of optimal β values clearly demonstrate optimal noise regularization is highly dependent on the acquisition time, injected dose, and
68Ga-labeled radiotracer.
Additionally, studies demonstrated the potential to reduce the acquisition time using BSREM while maintaining an image quality that is still comparable to OSEM reconstructions using longer acquisition times [
27,
29,
32]. Based on quantitative image analysis, up to a 75% reduction in acquisition time was achieved [
27], while another study reported a potential acquisition time reduction of up to 33% based on a visual assessment [
32]. Additionally, Svirydenka et al. demonstrated that the use of BSREM and a five-fold increase in the acquisition time resulted in a ten-fold reduction in the administered activity of
68Ga-PSMA in comparison to standard OSEM [
31], therefore enabling ultra-low activity examinations.
In addition to Q.Clear, United Imaging Healthcare introduces a BPL iterative reconstruction algorithm based on total variation regularized expectation maximization TVREM (HYPER Iterative) with a penalization factor between 0 and 1 to adjust the total variation penalization of voxels of corresponding neighborhoods. TVREM was used for PET phantom and patient scanning with
68Ga-PSMA (20 patients) [
35] and
68Ga-DOTATATE (17 patients) [
36]. As the penalization factor α increased, image noise [
35,
36], CR [
35], SUV
max [
35,
36], and TBR [
36] were reduced, while SNR increased [
36]. SUV
mean remained similar when the penalization factor increased [
35,
36]. Additionally, it was observed that the penalization factor α had a greater effect as the size of the lesions decreased. Specifically, for lesions with a diameter ranging between 10 and 20 mm, the SUV
max was decreased as the penalization factor α increased, while for lesions with a diameter greater than 20 mm, the SUV
max was almost identical [
35]. In addition to that, the relative differences in SNR and TBR were larger for lesions with a diameter lower than 10 mm than lesions with greater or equal [
36].
Quantitative image analysis and visual assessment by experts demonstrated that TVREM was capable of enhancing image quality more compared to OSEM but only for a certain range of values [
35,
36]. Yang et al. demonstrated that for penalization factor α ranging between 0.07 and 0.28 for
68Ga-PSMA, a reduction in the acquisition time of 33% was achieved based on quantitative image analysis and visual assessment from experts [
35]. Specifically, compared to OSEM, they reported improvements in the contrast up to 17%, the SUV
max up to 15%, and the image noise was reduced up to 32% [
35]. Liu et al. reported that TVREM improved the overall SUV
max, SNR, and TBR compared to OSEM [
36]. Additionally, based on visual assessment, they suggested that TVREM was capable of preserving the image quality for
68Ga-DOTATATE while achieving a reduction in the acquisition time by 33% for values between 0.14 and 0.35 [
36].
3.2.2. Deep Learning Approaches for Noise Reduction
The potential of deep learning approaches, both supervised and unsupervised, to reduce the noise in PET images has been demonstrated in the literature. For a supervised approach, low- and high-count PET images are available for each scan to train the network. Liu et al. implemented a supervised deep learning approach in which cross-tracer and cross-protocol learning were performed [
37]. They used PET images consisting of only 10% of the counts of the original single-bed
18F-FMISO (
n = 12), single-bed
18F-FDG (
n = 9), whole-body
18F-FDG (
n = 12), and whole-body
68Ga-DOTATATE (
n = 15) PET scans. Image quality enhancement was reported by comparison with the unenhanced, low-count PET images. Additionally, they reported the potential of employing neural networks trained based on
18F-FDG image data to reduce the image noise for other radiotracers such as
68Ga-DOTATATE and different scanning protocols [
37]. Deng et al. used low-count
68Ga-PSMA PET images (41 patients) and data augmentation to train a supervised deep learning model for denoising [
38]. Based on quantitative analysis and visual assessment, Deng et al. suggested a 50% count reduction as the optimal trade-off.
Additionally, unsupervised approaches were also investigated in the literature. In comparison to supervised approaches, training pairs are not required to train the unsupervised approaches, since there is no target output. Therefore, high-count PET images are unnecessary to train an unsupervised deep learning approach. Cui et al. investigated the potential of unsupervised deep learning [
39,
40]. An unsupervised neural network was validated using simulated
18F-FDG phantom data and evaluated using
68Ga-PRGD2 PET/CT (
n = 10) and
18F-FDG PET/MRI patient data (
n = 30) [
40]. Compared to the original PET images, the mean CNR was improved by approximately 53% for
68Ga-PRGD2 PET/CT and approximately 47% for
18F-FDG PET/MRI [
40]. However, they suggested radiotracer- and modality-independent effects since they reported no statistically significant differences between those improvements [
40]. In a later publication [
41], Cui et al. expanded their previous approach [
40] by incorporating a second unsupervised neural network, which used, as initial parameters, the pre-trained parameters of the first network. In that study [
41], the same dataset was used as in [
40]. This new approach improved the mean CNR by approximately 71% for
68Ga-PRGD2 PET/CT and by approximately 58% for
18F-FDG PET/MRI compared to the original PET images. In addition, compared to the previous approach [
40], it could better retain better structural image features [
41]. Meanwhile, SUV
max and SUV
mean values were comparable for both approaches [
41].