Background
Recently, new PET/CT scanners with long-axial field of view (LAFOV) and silicon photomultiplier (SiPM) detection systems have been introduced, such as the axial FOV of 194-cm PET/CT (uEXPLORER, United Imaging Healthcare Co), 106-cm PET/CT (Biograph Vision Quadra, Siemens Healthineers), and 64-cm PET/CT (PennPET Explorer) scanners [
1‐
4]. These LAFOV PET/CT scanners are characterised by increased sensitivity owing to their ability to collect more photons during scanning, allowing reduced tracer injection doses and shortened acquisition time.
Shorter acquisition durations are desirable for the comfort of patients, particularly those who are distressed, claustrophobic, have shortness of breath, are children, or require less dosage of anaesthetic; shorter acquisition durations is also cost-effective. One of the challenges in routine PET/CT is to deal with patients with dyspnoea in a recumbent position or severe pain following bone metastases. In such cases, maintaining diagnostic performance whilst achieving fast PET acquisitions would be especially beneficial.
However, shortening the acquisition time may result in increased noise, lower signal-to-noise ratio (SNR), and potentially unnecessary image artefacts [
5]. These factors may affect the quality and accuracy of the images, potentially compromising the diagnosis and treatment planning for the patient, especially for the detection of liver metastases or interference of physiological accumulation of
18F-FDG by the adjacent colon in colorectal cancer (CRC). Benign FDG uptake in the colon on PET/CT indicates physiological uptake, inflammation (such as inflammatory bowel disease), and benign lesions (such as benign polyps). This can affect the detection and diagnosis of CRCs, especially when the difference between tumour and colon benign uptake is not significant or when there is noise interference.
The advent of deep learning-based image filters has the potential to decrease noise and enhance image quality in short-term image acquisition, such as HYPER DLR launched by United Imaging Healthcare and licenced by the US Food and Drug Administration (FDA) 510(k) clearance. HYPER DLR, a deep learning-based algorithm for PET image filters, can effectively remove noise from images captured under low count rate conditions, significantly improving image quality. This technology can boost image SNRs by 42% and increase imaging speed [
6].
Xing et al. attempted to use HYPER DLR to reduce image noise and achieved good results, but did not evaluate the image quality with an acquisition time of less than 1 min [
6]. Some researchers have explored ultrafast PET acquisition and attempted to evaluate PET image quality within 1 min [
7‐
10]. However, image noise caused by short-term acquisition using a Gaussian filter has affected the diagnosis, with standard detectors covering an axial field of view [
11]. Although these studies on the ultrafast acquisition of PET/CT have revealed that acquisition speed has significantly improved, studies on the quality of images using a Gaussian filter still need to be completed. Further to this, previous studies have included a variety of diseases, but there is a lack of research on CRC, specifically in patients with liver metastasis. The effect of liver noise on the ability to detect metastatic tumours remains unknown.
Therefore, considering the 300-s OSEM reconstruction image with a 3 mm Gaussian smoothing filter as a standard image, we retrospectively compared the effects of the Gaussian and DL image filters on image quality, detection rate, and uptake rate of primary and metastatic CRCs in total-body PET/CT imaging at different acquisition durations (10, 20, 30, 60, and 120 s).
Methods
Study design and population
The local Institutional Review Board approved this retrospective study (No. KY2023-020-01) and waived the requirement for informed consent. From April 2022 to December 2022, 34 consecutive patients with CRCs were enrolled in this study.
Inclusion criteria
Patients were included in this study based on the following inclusion criteria: (i) no previous history of malignancies; (ii) the diagnosis was confirmed by histopathology; and (iii) only received symptomatic treatment and had no history of chemotherapy, radiotherapy, or surgical resection before the PET/CT scan.
Exclusion criteria
Patients were excluded for the following reasons: (i) incomplete image datasets and (ii) lack of a final histological diagnosis.
Patient preparation and PET/CT protocol
All patients fasted for more than 4 h before
18F-FDG injection according to the EANM procedure guidelines for tumour imaging (version 2.0) [
12]. All patients underwent implantation of 22 G indwelling intravenous catheters (Jierui Medical Product), followed by
18F-FDG manual administration with 2.96 MBq/kg. The patients were instructed to lie on the bed as calmly as possible. Imaging was started 60 ± 5 min after
18F-FDG injection. Image acquisition was performed using a total-body PET/CT scanner (uEXPLORER, United Imaging Healthcare, Shanghai, China) with an axial FOV of 194 cm. Additional file
1: Table 1 lists the parameters of the PET component of the PET/CT scanner. Low-dose CT was performed before PET for attenuation correction and anatomical localisation with a dose-modulation technique. Subsequently, total-body PET imaging was performed using a 3D list-mode with 300-s acquisition for one-bed position.
PET images were initially reconstructed with OSEM using data from the full 300-s acquisition. Images were post-processed using a 3 mm isotropic Gaussian smoothing filter. The necessary correction methods were applied, such as attenuation and scatter correction. Subsequently, the PET images were reconstructed using various acquisition times (10, 20, 30, 60, and 120 s) to simulate fast scans with both Gaussian and DL image filters. The parameters used in the OSEM reconstruction process included time of flight (TOF) and point spread function (PSF) modelling, three iterations, 20 subsets, 600 cm field of view, a matrix size of 192 × 192, a pixel size of 3.125 × 3.125 × 2.886 mm3, and a Gaussian post-filter of 3 mm FWHM. For the DL image filter process, the Gaussian post-filter was replaced, whereas all other reconstruction parameters were the same as described above.
Imaging analysis
All images were transferred to a workstation (uWS-MI:R002, United Imaging Healthcare) and reviewed in standard planes. Taking the 300-s image with a Gaussian filter image as a standard image, Gaussian and DL filtering images with five image datasets (10, 20, 30, 60, and 120 s) were included for comparison.
For qualitative analysis, the image quality of various time-point PET/CT datasets was evaluated visually by two nuclear physicians (Xiaochun Zhang and Taotao Sun) with over a decade of experience in PET/CT diagnosis. According to the widely used 5-point Likert scale, the image quality was scored, and the criteria were as follows: (i) very poor image quality and excessive noise (score 1); (ii) poor image quality and increased noise (score 2); (iii) fair image quality, similar to the regular image of daily practice (score 3); (iv) good image quality, superior to the regular image of daily practice (score 4); and (v) excellent image quality with minimal noise (score 5) [
13‐
17]. The two readers were blinded to the evaluation of the various time-point PET/CT dataset images and scored.
To eliminate intra-observer variability in the quantitative analysis, PET/CT images were quantitatively evaluated by a single nuclear physician with over a decade of experience in PET/CT diagnosis. Semi-automatic 3D delineation of the FDG-avid lesions was performed to cover the entire tumour. 3D isocontour volume of interest (VOI) based on 41% of the maximum standardised uptake value (SUVmax) thresholds was used and recommended by EANM guidelines [
12]. The tumour VOIs were obtained with the 300-s OSEM reconstruction with Gaussian (3 mm FWHM) filter and subsequently replicated and applied to other PET acquisition datasets of images. Mean standardised uptake value (SUVmean), maximum standardised uptake value (SUVmax), and peak standardised uptake value (SUVpeak) within a 1-cm
3 spherical volume were automatically generated.
According to the recommendations of PERCIST, hepatic
18F-FDG activity was assessed using a fixed 3-cm-diameter spherical VOI on the right lobe of the liver [
18,
19]. Additionally,
18F-FDG activity in the mediastinal blood pool was evaluated using a cylindrical VOI with a diameter of 1 cm and a long axis of 2 cm (parallel to the descending aorta) at the centre of the descending thoracic aorta. The SUVmean and standard deviation (SD) of the liver and mediastinal blood pool were recorded. The liver and mediastinal blood pool SNRs were calculated by dividing SUVmean by SD. The calculation formula used was as follows:
$${\text{Liver}}\;{\text{SNR}} = \frac{{{\text{SUVmean}}\;{\text{of}}\;{\text{Liver}} }}{{{\text{SD}}}}$$
$${\text{Mediastinal}}\;{\text{blood}}\;{\text{pool}}\;{\text{SNR}} = \frac{{{\text{SUVmean}}\;{\text{of}}\;{\text{Mediastinal}}\;{\text{blood}}\;{\text{pool}}}}{{{\text{SD}}}}$$
The tumour-to-background ratio (TBR) was calculated by dividing the SUVmax of the tumour by the SUVmean of the liver. The calculation formula used was as follows:
$${\text{TBR}} = \frac{{{\text{SUVmax}}\;{\text{of}}\;{\text{tumour}} }}{{{\text{SUVmean}}\;{\text{of}}\;{\text{Liver}} }}$$
To evaluate the detectability of the primary lesion of CRCs, the SUVmax of the adjacent proximal and distal bowel of the tumour was measured, and the tumour-to-adjacent bowel ratio (TAR) was calculated. The calculation formula used was as follows:
$${\text{TAR}} = \frac{{{\text{SUVmax}}\;{\text{of}}\;{\text{tumour}}}}{{{\text{SUVmax}}\;{\text{of}}\;{\text{the}}\;{\text{surrounding}}\;{\text{bowel}}\;{\text{of}}\;{\text{tumour}} }}$$
For patients with multiple liver metastases, the number of liver metastases was found on the 300-s image with a Gaussian filter image as a standard image reference, and the analysis focused on the largest and smallest lesions on the standard reconstruction. The liver metastases TBR and detection rate of liver metastases were calculated. The calculation formula used was as follows:
$${\text{Liver}}\;{\text{metastases}}\;{\text{TBR}} = \frac{{{\text{SUVmax}}\;{\text{of}}\;{\text{Liver}}\;{\text{metastases}}}}{{{\text{SUVmean}}\;{\text{of}}\;{\text{Liver}}}}$$
Statistical analysis
Continuous variables were presented as mean ± SD, and categorical variables were presented as frequencies and percentages. The weighted Kappa statistic was applied to evaluate inter-observer agreement for different acquisition durations of the PET/CT datasets image scores. The value of the agreement was categorised as follows: no agreement (κ < 0), slight agreement (0 ≤ κ < 0.2), fair agreement (0.21 ≤ κ < 0.4), moderate agreement (0.41 ≤ κ < 0.6), substantial agreement (0.61 ≤ κ < 0.8), and excellent agreement (0.81 ≤ κ ≤ 1). The Friedman’s test with post hoc comparisons using Bonferroni correction was used to compare differences in SNR, TBR, and tumour SUVs among various time-point PET/CT dataset images. Statistical analyses were performed using SPSS (v.26.0), GraphPad Prism (v.9.0.0), and MedCalc (v.19.0.7). A two-tailed probability value of < 0.05 was considered statistically significant.
Discussion
Our current study shows that DL image filter can significantly improve the image quality and SNR for low count data. Without affecting the quantitative evaluation of CRCs or liver metastases, the acquisition time of total-body PET/CT can be reduced to 20 s using DL image filter. For the visual qualitative evaluation of image quality, DL image filter significantly improved the image quality score of 20-, 30-, and 60-s acquisition time images compared with Gaussian filter (
P < 0.01). Visually, there was no difference in image quality between DL image filter images and 300-s images with Gaussian filter images (Additional file
8: Fig. 6). For the SNR of the liver and mediastinal blood pool, compared with Gaussian filter, DL image filter can increase the SNR of 20-s, 30-s, and 60-s datasets images by three times (Fig.
3A and 3B). Because of the 10-s datasets, whether Gaussian filter or DL image filter was too poor in SNR, it was not considered. Compared with the 300-s images with Gaussian filter images, the SNR of the 20- and 30-s images with DL image filter was similar.
When the acquisition time for PET/CT imaging is reduced, the image noise level tends to increase significantly. The SUVmax of CRC and liver metastatic lesions also tends to increase gradually. This observation applies to both images processed with a Gaussian filter and DL image filter. However, it is noted that the increase in SUVmax is relatively lower when using the DL image filter compared to the Gaussian filter. The detection of sub-centimetre liver metastases of CRCs is still a problem that puzzles PET/CT daily work. It is worth noting that DL image filter makes it difficult to detect sub-centimetre liver metastases from CRCs.
In this study, we compared the data from similar studies [
9]. The results are presented in Table
4. Compared with the current study, the SNR of the liver and mediastinal blood pool in Zhang et al.’s study is higher than that of our image data, whether it is 30- or 300-s image with a Gaussian filter. We speculate that this is caused by different FDG doses (3.7 vs. 2.96 MBq/kg).
Table 4
Quantitative comparison based on 300-s Gaussian filter image and comparison with relevant study
The current study | SNR | | | | | | | |
| Liver | 17.0 ± 4.4 | 5.6 ± 1.2 | P < 0.01 | 17.0 ± 4.4 | 18.7 ± 6.7 | 14.7 ± 6.2 | P > 0.05 |
| Aorta | 18.8 ± 6.9 | 6.2 ± 1.5 | P < 0.01 | 18.8 ± 6.9 | 19.4 ± 7.3 | 16.7 ± 6.9 | P > 0.05 |
| SUVmax | 21.6 ± 13.9 | 23.4 ± 13.9 | 0.063 | 21.6 ± 13.9 | 19.9 ± 12.9 | 20.4 ± 13.1 | P > 0.05 |
| TBR | 8.8 ± 4.9 | 9.4 ± 4.9 | 0.418 | 8.8 ± 4.9 | 8.1 ± 4.5 | 8.2 ± 4.6 | P > 0.05 |
Zhang et al | SNR | | | | | | | |
| Liver | 19.71 ± 5.58 | 9.03 ± 2.51 | P < 0.01 | | | | |
| Aorta | 16.60 ± 6.48 | 10.43 ± 3.1 | P < 0.01 | | | | |
| SUVmax | 13.94 ± 11.83 | 15.48 ± 14.19 | 0.003 | | | | |
| TBR | 4.43 ± 3.61 | 4.54 ± 3.80 | 0.411 | | | | |
Meanwhile, we also found that the SNR of the 20- and 30-s DL image filter images was higher than that of Zhang et al.’s 30-s Gaussian filter image [
9]. Additionally, the SNR of 30-s DL image filter image was similar to that of Zhang et al.’s 300 s with a Gaussian filter image [
9]. It has been demonstrated that the SNR of DL image filter images is better than that of Gaussian filter by a 30-s ultrafast acquisition by total-body PET/CT. And the 30-s DL image filter images was equivalent to the 300-s reconstruction images. The SUV and TBR of CRCs in this study were significantly higher than those in Zhang et al.’s study, presumably because of the different types of tumours included.
The increase in image noise due to fast acquisition may affect the detection and diagnosis of CRCs, especially when the difference between tumour and benign colon uptake is not significant or noise interference. Our study also evaluated the detectability of CRC lesions, compared the difference between CRC lesions and adjacent bowel benign FDG uptake, and used the SUVmax of lesions and adjacent bowel uptake. The shortened acquisition time led to deviations in SUVmax caused by noise.
Significantly, compared with the 300-s image with a Gaussian filter images, we have shortened the acquisition time and are more friendly to patients who cannot tolerate the conventional PET/CT acquisition time. In the routine management of CRCs, ultrafast acquisition may be a practical substitute for 300-s PET/CT acquisition.
Additionally, our study found that the SUVmax of the Gaussian filter image datasets acquired for more than 60 s was consistent with the results of the datasets acquired for 5 min (Fig.
4). Our results are consistent with Tan et al.’s research on assessing CRCs with data collected at 1–5 different minute time points using the same equipment [
20]. However, these results differ from those of Sher et al., who found that the SUVmax of images collected at 1.5 min differed from those collected at 5 min [
21]. This may be due to differences in equipment (long-axial FOV PET detectors and standard PET detectors) and the composition of the study cases, which are not focused on CRCs (mainly lymphoma and lung cancer).
Limitations of our study
The limitations of this study include its small sample size and retrospective design. One limitation of this study is that only one patient had a BMI greater than 30, defined as obesity. It is well known that obesity affects SNR and reduces the SNR of fast-acquisition PET images. This study also did not qualitatively and quantitatively evaluate the nearby lymph nodes of colorectal cancer, which will be the focus of the next step.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.