The elderly men and postmenopausal women had a high incidence rate of osteoporosis and related vertebral fracture (1). Vertebral fracture, especially thoracolumbar osteoporotic compression fracture, often occurs in the mid-thoracic (T7-8) and thoracolumbar spine (T12-L1) (2, 3). Bone mineral density (BMD) obtained from quantitative computed tomography (QCT) is a volumetric measure of vertebral trabecular bone with high sensitivity and accuracy for predicting bone strength and fracture risk (4–6). QCT not only reduces the influence of overlying ribcage (2) but also prevents severe spinal degeneration and vascular calcification without requiring the oral contrast agent and body position (5) compared with dual-energy X-ray absorptiometry (DXA). QCT is superior to DXA in BMD measurement for early screening of osteoporosis. However, a high level of radiation exposure delivered to patients with QCT limits its further clinical application (6). Recently, the combination of low-dose CT (LDCT) and lumbar QCT has been initiated by the China Health Big Data (China Biobank) project for opportunistic screening of osteoporosis and lung cancer simultaneously in terms of reducing radiation dose, repeated scan, patient time, and additional costs. Wu et al. (5) described the study protocol of the combination of QCT with LDCT. Inherently, Cheng et al. (7) conducted a multicenter population-based cohort study with QCT to determine the prevalence of osteoporosis in China.

Unfortunately, image noise increased obviously after reducing radiation dose, while image quality decreased significantly, particularly in the spine (5), contributing to an inevitable decrease in diagnostic performance. An iterative reconstruction (IR) algorithm is introduced to reduce image noise and preserve image quality between radiation risk and diagnostic performance (8, 9). But many IR algorithms can change the magnitude of the image noise and texture details and may cause an adverse impact on the detection of low-contrast lesions, particularly at high strength levels (10–12).

Currently, a new-generation deep-learning image reconstruction (DLIR) (TrueFidelity, GE Healthcare) was proposed to improve the CT image quality. It utilizes deep neural networks that consist of layers of mathematical equations, with millions of connections and parameters to generate CT images, and is designed with a fast reconstruction speed for routine CT use, even in acute care settings. And it consists of three selectable reconstruction strength levels (low, medium, and high) to control the amount of noise reduction corresponding to clinical applications and radiologist preference (13).

To assess the image quality of LDCT, accurate BMD measurement, and the performance of DLIR for image quality at ultralow-dose level, Li et al. (14) systemically evaluated the physical image quality on Catphan 500 phantom. Results indicated that the CT number linearity was unbiasedly contributing to accurate BMD quantification. DLIR performed better than iterative model reconstruction (IMR, level 2) at 0.25 and 0.75 mGy, but they didn’t evaluate the accuracy of BMD value on European Spine Phantom (ESP). Therefore, on the basis of Li et al.’s experiment, our study aimed to evaluate CT image quality and accurate BMD measurement on the Catphan 500 phantom and ESP and patient study using DLIR algorithm in comparison to 50% adaptive statistical iterative reconstruction-veo (ASiR-V 50%) and filtered back projection (FBP) reconstruction algorithms.

This prospective study was strictly adhered to HIPAA Privacy Rule and approved by the ethics committee of the First Affiliated Hospital of Zhengzhou University and Beijing Jishuitan Hospital. The China Biobank project is a multicenter cohort study and has been registered with the US clinical trials database (https://clinicaltrials.gov/ct2/show/NCT03699228; trial identifier: NCT03699228). Our hospital is one of the collaborating hospitals and provided the patient cohort for this study. The informed consent of the patients was all obtained.

Data acquisitions were obtained from Catphan 500 phantom (Phantom Laboratory, Salem, NY, USA) and ESP (No. 145, Germany ORM company), as well as patients on Revolution CT (GE Healthcare, WI, USA) from April 2020 to June 2021. The weekly air calibration and monthly QA were performed by qualified technologists before data acquisitions and BMD measurement throughout the whole study using the Model 3 synchronous QA phantom. To reduce the uncertainty of measurements, data acquisitions were scanned 10 consecutive times separately on Catphan 500 and ESP without repositioning.

In our study, we systematically evaluated the image quality, accurate BMD measurement, and clinical applicability of QCT with DLIR based on multi-phantom and patient studies. Results indicated great clinical importance without requiring any additional equipment and patient time, repeated CT scan, radiation dose, and additional costs. To our knowledge, it is the first systemic study to research the application of BMD measurements at an ultralow-dose level. QCT can be utilized for further opportunistic screening of osteoporosis, osteoporotic fracture, or other clinical applications (e.g., health check-ups) in China or worldwide countries accessing to CT easily than DXA (7).

Our results are consistent with Li et al. (15) findings on Catphan 500. For three levels of DLIR, MTF value at 50%MTF was about 4.50lp/cm, better than those for FBP (4.12 lp/cm) and ASiR-V50% (4.00 lp/cm). The 2 or 3 mm low-contrast object was clearly resolved at a 0.5% contrast level or at FBP and ASiR-V50%. Abdullah et al. (16) reported that the 50%MTF value and smallest size of objects were about 0.41 lp/cm and 3mm with ASiR-V (level: 40% and 60%), slightly lower than 4.50lp/cm and 2mm with DLIR. It showed an obviously lower NPS peak frequency and noise level, and a shift towards a lower spatial frequency in NNPS curve. As the strength level increased, the peak and spatial frequency of NPS curves with DLIR were decreased, which is consistent with a study reported by Greffier et al. (26). DLIR has been developed to reduce radiation dose and maintain image quality without changing the image texture or affecting the anatomical and pathological structures (13). And it can decrease the low-frequency noise component to improve low-contrast detectability for soft tissues ranging from 50 to 200 HU in abdominal CT (27), while maintaining the high-contrast spatial resolution of detailed structures, such as sharp edges and vessel boundaries at a low-dose level.

For image analysis in patients, DLIR-M and DLIR-H were scored better than ASiR-V50% in image noise, image contrast, small structure visibility, image sharpness, and artifacts. As radiation dose and strength level increased, image noise significantly decreased, CNR obviously increased, whilst CT value showed no significant difference (P>0.05). Results indicated that DLIR had better overall image quality than ASiR-V50%. Our finding was in accordance with Singh et al. (28) and Kim et al. (29)’s study that both obtained with relatively small sample sizes, but revealed a better significance due to the large patient cohort. Several studies suggested that DLIR was scored significantly better in overall image quality than different strengths of ASiR-V (level: 30%, 40%, and 50%) (24, 29) and comparable to ASiR-V (level: 70%, 100%) (30, 31).

Three HA inserts of 50.2-199.2 mg/cm³ provided a range of trabecular BMD mimicking the physiological range of BMD seen in all age groups (32). The relative measurement error of L1, L2, and L3 was 9.84%, 4.08%, and 2.60%, respectively. Coefficients of variance for the L1, L2, and L3 HA inserts were 1.51%, 1.41%, and 1.18%. Those all falling within the range of 4-15% and meeting the clinical BMD measurement requirements (4, 32, 33). The largest and smallest deviations were found in L3 and L1, respectively. As the BMD value decreased, the relative measurement error increased significantly; especially with BMD less than 100.2 mg/cm³, thus more attention should be paid to osteoporosis patients when evaluating the risk of osteoporotic fractures. Wu et al. (4) investigated the repeatability and accuracy of QCT measurement of BMD by low-mAs with iterative model reconstruction (IMR) algorithm based on phantom level and showed the maximum deviation of accuracy was 11% for L1, 4% for L2, and 6% for L3. In contrast, our study demonstrated that the accuracy of BMD at L1 and L3 was improved with DLIR in comparison to IMR (2), indicating that DLIR may potentially improve the low-contrast detectability and maintain the high-contrast spatial resolution. However, further studies should be implemented to verify whether DLIR can makes the images more homogeneous in terms of CT numbers. Consistent with our findings, Wang et al. (6) observed an excellent accuracy with 3 HA inserts ranging from 3.7% to 5.9%. Zhao et al. (16) found that the mean trabecular BMD measurement of 3 HA inserts were 2.4%, 2.1%, and 0.5% at L1, L2, and L3 for forty different systems on ESP, indicating a smaller measurement error than our study.

For patients aged over 50 years, the prevalence rate of osteoporosis was 24.67% in women and 11.79% in men, and it was comparable to 29.1% in women but more than twice in men by DXA, and similar to 29.0% in women and 13.5% in men by QCT reported by Cheng et al. (7). The prevalence rate of osteoporotic fracture was 8.67% in women and 5.29% in men, which was significantly lower than 17.3% in women and 17% in men for more than 14000 subjects in Shanghai conducted by Gao et al. (34). Conversely, a study in Norway enrolled 2887 participants demonstrated a higher prevalence rate of vertebral fracture 11.8% in women and 13.8% in men (35). The difference in osteoporotic fracture between DXA and QCT may be attributed to the patient cohort mostly obtained from the health check-up participants for osteoporosis screening, thus further studies should be performed to assess the fracture risk of QCT in multiple participants.

There are some limitations to be highlighted. Firstly, the results acquired with QCT should be further compared with DXA corresponding to the prevalence of osteoporosis. Secondly, a longitudinal study should be further performed to verify the clinical utility of DLIR algorithms in osteoporosis screening. Thirdly, we didn’t evaluate the risk factors of osteoporosis, such as age, BMI, smoking, and fragility fracture history.

In conclusion, image quality with DLIR was high-qualified without affecting the accuracy of BMD measurement. It may provide a great clinical utility in osteoporosis screening.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the Ethics Committees of First Affiliated Hospital of Zhengzhou University. The patients/participants provided their written informed consent to participate in this study. Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

YL, YJ, and YW designed the study. YL and YJ performed the data analysis. YL researched the related literatures. All authors contributed the data collection, measurements, and interpretation. YL wrote the manuscript and all authors reviewed the manuscript.

This study is supported by the National Natural Science Foundation of China (grant no. U1504821).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.