DOI: https://doi.org/10.21203/rs.3.rs-1964028/v1
Purpose: To investigate the feasibility of ultra-low dose CT (ULDCT) reconstructed with the artificial intelligence iterative reconstruction (AIIR) algorithm in total-body PET/CT imaging.
Methods: The study included both the phantom and clinical part. An anthropomorphic phantom underwent CT imaging with ULDCT (10mAs) and standard-dose CT (SDCT) (120mAs) respectively. ULDCT was reconstructed with AIIR and hybrid iterative reconstruction (HIR) (expressed as ULDCT-AIIRphantom and ULDCT-HIRphantom) respectively, and SDCT was reconstructed with HIR (SDCT-HIRphantom) as control. In the clinical part, 52 patients with malignant tumours underwent the total-body PET/CT scann. ULDCT with AIIR (ULDCT-AIIR) and HIR (ULDCT-HIR) respectively was reconstructed for PET attenuation correction, followed by the SDCT reconstructed with HIR (SDCT-HIR) for anatomical location. PET/CT images quality were qualitatively assessed by two readers. The CTmean as well as the CT standard deviation (CTsd), SUVmax, SUVmean, and the SUV standard deviation (SUVsd) were recorded. The signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were calculated and compared.
Results: The image quality of ULDCT-HIRphantom was inferior to the SDCT-HIRphantom, but no significant difference was found between the ULDCT-AIIRphantom and SDCT-HIRphantom. The subjective score of ULDCT-AIIR in the neck, throax and lower limb was equivalent to that of SDCT-HIR. Besides the brain and lower limb, the CTmean of the lesion and other regiongs in ULDCT-AIIR didn’t fluctuate too much, while the CTsd of ULDCT-AIIR was lower than that of SDCT-HIR. Except for the brain, the CNR of ULDCT-AIIR was the same as the SDCT-HIR, but the SNR was higher. The SUVmax, SUVmean and SUVsd of PET-AIIR had a slightly change in the whole body. For the lesions, the SUVmax, SUVsd and TBR showed no significant difference between PET-AIIR and PET-HIR.
Conclusion: The SDCT-HIR was not replaced by the ULDCT-AIIR, but the AIIR algorithm decreased the image noise and increased the SNR, which can be implemented under special circumstance in PET/CT examination.
PET/CT noninvasively provide both physiological and anatomical information in vivo in a single gantry, which has become a powerful imaging modality in the diagnosis, staging, prognosis, treatment planning, and therapy response assessment [1–3]. Recently, the concerns about radiation-exposure-induced carcinogenesis in PET/CT examination have raised, especially for the patients undergoing the repeat scans for treatment evaluation. The trade-off between the diagnostic need and the radiation exposure in the PET/CT examination should be balanced, particularly for pediatric patients [4, 5]. With the advance of the total-body PET/CT scanners with the improved sensitivity, studies have proven to have the potential of fast and ultra-low-dose PET imaging in the clinical practise [6–8].
The CT protocols used in PET/CT can be divided into three types according to the different applications: diagnosis, anatomical localization and only for PET attenuation correction. CT dose was the major contributor to the effective activity in PET/CT procedures [9]. CT dose reduction techniques such as tube current modulation, tube voltage reduction, pitch increasement and noise reduction filters, and new reconstruction algorithms were proposed and implemented in the past years [10] [11]. Among them, adaptive statistical iterative reconstruction algorithm, such as HIR, were implemented for dose reduction and serves as a regular reconstruction algorithm in the daily clinical. It consists of a scanner model, a statistical noise model, and projection noise estimation, which optimizes raw data in the projection domains through iteration loops [12]. The HIR suppresses the image noise, but may lead to “plastic” images and degrade diagnostic accuracy [13]. The AIIR is a deep-learning based reconstruction algorithm that trains the deep learning network with a big dataset of CT images [14]. This newly developed reconstruction algorithm is capable of denoising low-dose images while maintaining the image quality compared to the SDCT images [14–16]. Thus, the radiation exposure could be theoretically further reduced with the AIIR, but few studies on the ULDCT scans has been conducted.
This study aims to investigate the potential of AIIR in dose-reduction in the total-body PET/CT system retrospectively, and compare the image quality between the AIIR-reconstructed ULDCT and HIR-reconstructed SDCT images.
An anthropomorphic phantom, Kyoto Kagaku PBU-60 (Kyoto Kagaku Co., Kyoto, Japan), was assembled according to the instruction manual, and placed in the head-first position. The phantom was scanned in two dose levels, which were consistence with the clinical acquisition protocols for ULDCT and SDCT.
The CT images were reconstructed with AIIR and HIR with the convolution kernel (level 5), B-Soft-B (United Imaging Healthcare, China) for ULDCT: ULDCT-HIRphantom and ULDCT-AIIRphantom, and SDCT-HIRphantom as control.
Ten ROIs were draw on CT images at the white matter, grey matter, CSF, thoracic vertebra, heart, air in the lung, liver, kidney, muscle, and femur of the anthropomorphic phantom. The image signal (CTmean) and image noise (CTsd) of the ROIs were statistically assessed. The SNR and CNR were calculated for each ROI to compare the image quality for each group defined as follows:\(SNR= \frac{{CT}_{mean, ROI}}{{CTsd}_{, ROI}}\) [17], \(CNR= \frac{{CT}_{mean,ROI}-{CT}_{mean, Ref}}{({CT}_{sd,ROI}+{CT}_{sd,Ref})/2}\) [18] with muscle as reference.
This retrospective study was approved by the Institutional Review Board of Zhongshan Hospital, Fudan University and written informed consent was signed for the use of the data from all the subjects. 52 patients (mean age 59.48 ± 13.62; 29 male, range 41–86 years; 23 female, range 21–86 years) were randomly enrolled in this study from October to November 2021 (Table 1). The exclusion criteria included: no uptake of FDG in the primary lesions, blood glucose level > 11.0 mmol/L, an FDG uptake time of more than 70 min, and the images severely affected by the motion artifact.
| Characteristic | Value | |
|---|---|---|
| Age (year, range) | 59.48 ± 13.62 (21–86) | |
| Sex | ||
| Male (%) | 29 (55.8%) | |
| Female (%) | 23 (44.2%) | |
| BMI | 23.37 ± 2.74 (18.01–30.26) | |
| Blood glucose (mmol/L) | 5.90 ± 1.22 (3.4–10.4) | |
| Injected activity (MBq/kg) | 120.17 ± 16.53 (76.22-157.99) | |
| Location of primary tumors | ||
| Head and neck | 2 (3.85%) | |
| thorax | 16 (30.77%) | |
| Abdomen and pelvic | 33 (63.46%) | |
| Lower limb | 1 (1.92%) | |
The examinations were performed on the total-body PET/CT scanner, uEXPLORER (United Imaging Healthcare), equipped with a 128-slice CT. The patient lied in supine position and underwent CT scout for the correction of the positioning, then followed by an ACCT for the PET attenuation correction which was used as ULDCT in this study. After the acquisition of the PET data, the range of SDCT for anatomical location was scanned from the top of the skull to the middle femur. The scanning protocol of the total-body ULDCT was set at the tube voltage of 120 kV, tube current of 10 mAs, gantry rotation time of 0.5 s, beam pitch of 1.0125, slice thickness of 1.0mm, field of view of 500 mm. The acquisition protocol of SDCT was tube voltage 120 kV, auto modulation tube current with reference 120 mAs, gantry rotation time 0.5 s, beam pitch 0.9875, slice thickness 1.0mm, field of view 500 mm.
The ULDCT images were reconstructed with AIIR and HIR using the convolution kernel (B-Soft-B, United Imaging Healthcare), expressed as: ULDCT-AIIR and ULDCT-HIR. The SDCT images were reconstructed by HIR with convolution kernel and served as reference, noted as SDCT-HIR. PET reconstructions were performed using the OSEM algorithm with the following parameters: TOF and PSF modelling, 3 iterations, and 20 subsets, matrix of 192 × 192, slice thickness of 1.443 mm, and a full width at half maximum of the Gaussian filter function of 3 mm. The attenuation of PET images was corrected by the ULDCT-AIIR and ULDCT-HIR respectively (expressed as PET-AIIR and PET-HIR).
Two nuclear medicine physicians with 5 and 7 years of experience were blinded to the dose and reconstruction methods, and viewed the images using standard window settings. The images were graded on the scale of 1–5 in the aspect of the image noise, conspicuity of anatomical structures, and diagnostic confidence for all the images from each subgroup. The conspicuity of anatomical structures includes the visualization of the brain, neck, thorax, abdomen, pelvic and lower limb. The subjective score was evaluated according [19]. 1: non-diagnostic, unacceptable quality. impossible to detect important anatomic structures (clearly evaluate their margin or internal characteristics) due to high noise or severe artifacts. 2: poor image quality. possible to detect crucial anatomic structures (difficult to clearly evaluate their margin or internal characteristics) with distinctly increased noise or considerable artifacts. 3: acceptable image quality. depiction of the anatomic structures (margin or internal characteristics can be detected) with noticeable noise or some distinct artifacts, may obscure very subtle details. 4: good image quality. depiction of all anatomic structures (easier to evaluate their margin or internal characteristics) with slightly increased noise, above average image quality with the noise and artifacts not affecting diagnostic value. 5: excellent image quality. depiction of all anatomic structures with fine details (significantly easier to evaluate their margin or internal characteristics, without any indistinct findings), free of artifacts and with imperceptible noise.
For all three groups in the clinical study, three ROIs were drawn at the brain, including white matter, grey matter, CSF; two ROIs at the neck, including thyroid, neck muscles; three ROIs were drawn on the thorax, including lung, mediastinum, back muscle; five were drawn at the abdomen, including liver, lumbar muscle, first lumbar spine, sigmoid colon, gluteus maximus; two were in the lower limb, including muscles, lower limb bone; the primary lesion was also delineated from each patient. Objective image quality for the CT images was also assessed with the CTmean, CTsd, SNR and CNR. The muscles were considered as reference in each section of the body for the calculation of CNR. The SUVmax, SUVmean and its standard deviation (SUVsd) were measured for PET images. TBR was calculated as follow: \(TBR= \frac{{SUV}_{max, lesion}}{{SUV}_{mean, liver}}\).
Statistical analysis was conducted using R software version 4.1.2 (R Core Team) and GraphPad Prism 6.0 (GraphPad Software Inc.). The results of objective assessment were expressed as mean \(\pm\) SD. The significance of difference in the objective evaluation between each subgroup was measured by the Wilcoxon signed rank with Bonferroni correction to minimize the possibility of significance brought by multiple statistical testing. A p value < 0.05 was considered statistically significant. The inter-observer agreement was analyzed by the Cohen’s kappa test where the k value in the range of 0.00 to 0.20 indicated poor agreement; k value in the range of 0.21–0.40 indicated fair agreement; k value in the range of 0.41–0.60 indicated moderate agreement; k value in the range of 0.61–0.80 indicated good agreement, k value in the range of 0.81-1.00 indicated excellent agreement. Pearson correlation coefficient (r) was used to measure the linear relationship between the SUVmax, SUVmean and SUVsd of PET-HIR and PET-AIIR. The r between 0.8-1.0 indicated very strong linear relationship.
The CTmean of ULDCT-AIIRphantom was equivalent to that of SDCT-HIRphantom (p = 0.761), while the ULDCT-HIRphantom was significantly different from that of SDCT-HIRphantom (p = 0.049). No statistically difference was found between the image noise of SDCT-HIRphantom and ULDCT-AIIRphantom (p = 0.241), while the CTsd of ULDCT-HIRphantom is significantly higher than that of SDCT-HIRphantom (p < 0.001). There is no significant difference between SDCT-HIRphantom and ULDCT-AIIRphantom regarding to SNR and CNR (p = 0.058, p = 0.241). However, the SNR and CNR of ULDCT-HIRphantom were inferior to the SDCT-HIRphantom (both p < 0.001) (Fig. 1). As shown in Fig. 2, the ULDCT-HIRphantom was eliminated from the comparition regarding the CT image quality in the subsequent clinical experiment due to its heavy noise.
The subjective scores of two CT images showed good agreement between two nuclear medicine physicians (k = 0.755). Compared with the SDCT-HIR, ULDCT-AIIR achieved the equivalent image quality in the neck, thorax and lower limb. While the subjective scores dropped in the brain (3.81 ± 0.26 to 2.23 ± 0.33) and abdomen (3.86 ± 0.27 to 2.17 ± 0.34) in ULDCT-AIIR (Table 2). As shown in Fig. 3, the anatomical details in ULDCT-AIIR images were more obscure than that in SDCT-HIR images.
| Index | Subjective scores HIR | Subjective scores AIIR | Weighted Kappa | ||
|---|---|---|---|---|---|
| Overall | 3.81 ± 0.41 | 3.18 ± 0.90 | 0.755 | ||
| Brain | 3.81 ± 0.26 | 2.23 ± 0.33 | 0.554 | ||
| Neck | 3.43 ± 0.47 | 3.65 ± 0.47 | 0.829 | ||
| Thorax | 4.05 ± 0.42 | 4.03 ± 0.44 | 0.743 | ||
| Abdomen | 3.86 ± 0.27 | 2.17 ± 0.34 | 0.744 | ||
| Lower limb | 3.89 ± 0.29 | 3.80 ± 0.37 | 0.677 | ||
As shown in Fig. 4, the CTmean of the lesion, neck, thorax, abdomen and pelvic in ULDCT-AIIR was equivalent to that in SDCT-HIR (all p > 0.05), while significant difference was found in the grey matter, cerebrospinal fluid, and lower limb skeleton (p < 0.05). Other than in brain, thyroid, lung, sigmoid colon, and lower limb skeleton, CTsd of the ULDCT-AIIR images was lower than that of the SDCT-HIR images. The CTsd of ULDCT decreased in the neck muscle (-10.80%), back muscle (-17.08%), mediastinum (-15.77%), liver (-26.74%), lumbar muscle (-20.28%,), first lumbar spine (-27.84%), gluteus maximus (-21.02%), lower limb muscles (-10.63%), and lesion (-12.09%). Besides the brain, neck, lung, sigmoid colon and lower limb, the SNR of the ULDCT-AIIR was higher than that of the SDCT-HIR (p < 0.001). The CNR of the ULDCT-AIIR in brain was statistically difference with SDCT-HIR (p < 0.001), but no significant difference was found in other regions.
Compared with the PET-HIR, the SUVmax, SUVmean and SUVsd of PET-AIIR slightly changed in the whole body, although significant differences were found in some anatomical locations. The correlation analysis showed a high degree of consistency of these parameters between the PET-HIR and PET-AIIR (r > 0.99, supplementary material). For the lesions, the SUVmax, SUVsd and TBR didn’t fluctuate significantly between PET-AIIR and PET-HIR (Table 3). Besides the cerebrospinal fluid, the SUVsd between the PET-AIIR and PET-HIR showed no statistical difference.
| Index | HIR | AIIR | Change Rate | P-value |
|---|---|---|---|---|
| Lesions | ||||
| CTmean | 35.08 ± 14.07 | 41.55 ± 19.33 | 25.87% | 0.523 |
| CTsd | 14.61 ± 6.04 | 11.63 ± 5.18 | -12.09% | 0.028 |
| SUVmax | 6.08 ± 3.50 | 6.06 ± 3.49 | -0.27% | 1 |
| SUVmean | 5.38 ± 3.05 | 5.38 ± 3.07 | -0.15% | 1 |
| SUVsd | 0.55 ± 0.44 | 0.55 ± 0.44 | -0.82% | 1 |
| TBR | 2.53 ± 1.60 | 2.54 ± 1.60 | 0.26% | 1 |
| SNR | 2.79 ± 1.41 | 4.30 ± 2.52 | 71.91% | 0.003 |
The reduction of radiation dose from PET-CT studies is vital for pediatric patients and the patients who often undergo multiple follow-up studies [20]. It can be achieved by reducing the amount of the injected radiopharmaceutical and optimizing the CT dose in the PET/CT examination [21, 22]. Optimization of X-ray tube current and tube voltage can reduce radiation exposure, and the advanced reconstruction algorithms [23, 24] can achieve the comparable image quality with lower dose. When CT is only used for PET attenuation correction, the range of exposure settings is extremely wide [25]. Careful consideration of the acquisition parameters is required to achieve low-dose image with acceptable quality. PET/CT scanner has imaging protocols with a default value of tube current, with an option to change it to suit the patient study and the radiation doses should be kept as low as reasonably achievable. Perfect CT images, which are more appealing to our eyes, however, may not add extra diagnostic information and are obtained at the cost of giving unnecessary radiation burden to the patient [26].
The ULDCT was used as attenuation correction for total-body PET/CT in our clinic institution, which added additive radiation compared to the conventional PET/CT. The ULDCT reconstructed with AIIR algorithm was evaluated in this study to explore the possibility of further dose reduction in the total-body PET/CT examination. The phantom study demonstrated that the AIIR reconstruction algorithm in ULDCTphantom didn’t affect the CTmean value, while the AIIR algorithm reduced the noise and improved the SNR of ULDCT images, and had the same level compared with SDCT-HIRphantom. This phantom study confirmed the feasibility of AIIR algorithm in ULDCT. In the clinical study, the image quality of ULDCT-AIIR in the neck, throax and lower limb was equivalent to the SDCT-HIR. In the brain and the abdomen, the details in ULDCT-AIIR were more obscure than these in SDCT-HIR, that the hands of the patient were placed on the sides of the body in this study. Thus, it was difficult for X-ray to penetrate the skull and bilateral upper limbs with high density. Obviously, the serious streaking artifact was caused by the photon starvation [27]. The attenuation was greatest and insufficient photons reached the detectors, which produced very noisy projections at the tube angulations and resulted the horizontal artifacts. To some extent, this may affect the CTmean and SNR of these parts. The streaking artifacts will be alleviated with sufficient tube current, but the patients will receive unnecessary dose of radiation in this way [28]. Noted that CTsd in some anatomical parts in ULDCT with AIIR was lower than that with HIR-reconstructed SDCT images. Besides the brain, the CNR with ULDCT-AIIR was equivalent to that with SDCT-HIR. This indicated that the noise of ULDCT images was greatly decreased by the AIIR, and the CNR was kept as high as possible. The difference between the phantom and clinic study was that the phantom was a relatively ideal model, which derived from the standard patient morphology, biokinetics and organ sizes [29] and cannot completely represent the precise tissue structure of human body.
Theoretically, the tube current reduction should have limited effect on the SUVmean or SUVmax, that the attenuation correction CT often requires less tube current to generate a low-resolution attenuation map [24]. Kumar et al found that the SUVmean varies from 1.2 to 1.4 with the changing of tube current, but the change in tube current didn’t change much in the SUVsd, and Fahey et al found that CT noise yielded a 2% variation in SUVsd [26, 30]. However, the attenuation correction CT image for PET was the ULDCT-AIIR and ULDCT-HIR respectively, which was only related on the reconstruction algorithm. We found that the SUVmean in the brain and neck was increased in PET-AIIR group but decreased in the thorax and abdomen comparing to that of the PET-HIR group, though these slight fluctuations had no practically clinical significance. Except for the cerebrospinal fluid, the SUVsd showed no significant change between two groups, meaning that the noise of PET images does not change with the reconstruction algorithm. For the lesions, TBR between two groups also showed no significant difference, the reconstruction algorithm was not its contributory factor. This finding was similar to the previous published study, which reported that reconstruction algorithm had no change in SUV [31].
Some limitations exist in our study. Firstly, the patients were not grouped by the weight, but the dose of CT was unified. The effective activity in body could be affected by the individual differences [17, 19]. Secondly, the positioning in our study have affected the image quality and the measurement of lesion. The high-density and the low-density strip artifact was caused by the passage of high-energy X photons and the starvation of photons, respectively. Further studies may be conducted to determine the suitable positioning for total-body PET/CT. Thirdly, no pediatric patients were included in this study, which is a targeted population for radiation dose-reduction strategies. The peculiarities of acquisition parameters for pediatric patients limited the type of patients included.
The image quality of ULDCT-AIIR was not equivalent to that of SDCT-HIR in some anatomical parts, that the CTmean and SNR were affected while CNR and TBR were unaffected. The minor changes of SUVmean and SUVsd had no significance in clinic. Overall, the image quality of ULDCT-AIIR was not superior to SDCT-HIR, but the AIIR algorithm decreased the image noise and relatively increased the SNR, which can be implemented under special circumstance in PET/CT examination.
PET Positron Emission Tomography
CT Computed Tomography
ULDCT ultra-low-dose CT
SDCT standard-dose CT
HIR hybrid iterative reconstruction
AIIR artificial intelligence iterative reconstruction
SUV Standardized Uptake Value
SNR Signal-to-Noise Ratio
CNR contrast-to-noise ratio
CSF cerebrospinal fluid
ROI regions-of-interest
FDG Fluorodeoxyglucose
ACCT Attenuation-Correction CT
OSEM ordered subset expectation maximization
TOF time of flight
PSF point spread function
TBR tumour-to-background ratio
Ethics approval and consent to participate All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. This retrospective study was approved by the Institutional Review Board of Zhongshan Hospital, Fudan University and written informed consent was signed for the use of the data from all the subjects.
Consent for publication All authors discussed and approved the final manuscript for submission.
Availability of data and material Availability of data and material The data that support the findings of this study are available from the corresponding author upon reasonable request.
Competing interests The authors declare that they have no competing interests.
Funding This study is supported by the National Science Foundation for Scholars of China (No. 81871407) and Shanghai Municipal Key Clinical Specialty (No. shslczdzk03401), Three-year Action Plan of Clinical Skills and Innovation of Shanghai Hospital Development Center (No. SHDC2020CR3079B), the Shanghai Science and Technology Committee (No. 20DZ2201800), Three-year Action Plan for the fifth round of public health system construction in Shanghai (No. GWV-10.1-XK10), and Next Generation Information Infrastructure Construction Project (No. 201901014). No other potential conflicts of interest relevant to this article exist.
Authors' contributions Yan Hu and Hongcheng Shi contributed to the study design. Yan Hu, Jingyi Wang and Xinlan Yang contributed to the data processing and analysis. Haojun Yu and Zhe Zheng contributed to the image acquisition and processing. Yan Hu and Hongcheng Shi contributed to the manuscript drafting. Hongcheng Shi contributed to the article revising. All authors discussed and approved the final manuscript.
Acknowledgements Not applicable.