Ultra-low-dose hepatic computed tomography with a novel real-time deep learning-based noise reduction algorithm: a prospective cross-sectional analysis of image quality and lesion detection

Introduction

Contrast-enhanced computed tomography (CT) has become an indispensable imaging modality for tumor staging, assessing the therapeutic response, detecting recurrence, and identifying new lesions and/or distant metastasis (1-3). However, evaluating low-contrast lesions remains a significant challenge in medical imaging, particularly for liver metastasis. Reducing the radiation dose in CT scans is beneficial but often leads to increased image noise, which may degrade image quality and limit diagnostic performance, especially for small or low-contrast lesions (4,5). Iterative reconstruction (IR) algorithms have been developed to reduce image noise while attempting to preserve image quality. Despite their advantages, several studies have shown that high-strength IR can alter both image noise and the texture details, potentially impairing the detection of low-contrast lesions (6,7).

Recently, deep-learning image reconstruction (DLIR), leveraging convolutional neural networks, has gained attention as a promising alternative to IR and sought to address its limitations. Studies applying DLIR to chest, upper abdomen, coronary, and brain CT have demonstrated a reduction in the effective dose (ED), while ensuring image quality and enhancing lesion diagnosis capabilities (8-12). Previous research on first-generation DLIR focused on focal lesions, and reported improved CT image quality with a moderate (50%) reduction in the radiation dose, but found that a 70% dose reduction was insufficient for detecting small, low-contrast hepatic metastases (13). To date, DLIR applications in ultra-low-dose (ULD) scenarios remain limited. Two recent studies reported a 76% reduction in radiation dose while maintaining diagnostic image quality (14,15). However, these studies primarily focused on image quality improvement, without assessing diagnostic lesion detection, particularly for low-contrast lesions. There remains a significant discrepancy in diagnostic standards between image qualities for low- and high-contrast abdominal lesions (16).

Therefore, this prospective study sought to apply a novel real-time DLIR algorithm [ClearInfinity (CI), Neusoft Medical Systems, Shenyang, China] in ULD abdominal CT, and more specifically to assess its ability to preserve the detection of hepatic lesions of varying sizes and contrasts. We present this article in accordance with the STROBE reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-365/rc).

Methods

This prospective study was approved by the Human Research Ethics Committee of The First Affiliated Hospital of Zhengzhou University (No. 2022-KY-0752-001), and was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. All patients provided written informed consent before participation.

Participants and study design

The sample-size calculation was performed using PASS 15 (Power Analysis and Sample Size Software, Version 15), with image noise identified as the primary image quality indicator requiring the largest sample size based on pre-experimental data. A repeated-measures analysis of variance was applied for the analysis. The power analysis indicated that 63 participants would provide at least 95% power with a one-side type I error rate of 10%. Between March 2023 and October 2023, a total of 96 patients with suspected hepatic lesions were enrolled for contrast-enhanced abdominal CT scans following the standard protocol of The First Affiliated Hospital of Zhengzhou University. The inclusion and exclusion criteria are detailed in Figure 1.

Figure 1 Flowchart of the study design. Group A: standard-dose protocol with 50% CV; group B1: ultra-low-dose protocol with 50% CV; group B2: ultra-low-dose protocol with 50% CI. CI, ClearInfinity; CT, computed tomography; CV, ClearView.

Of the 96 patients, 31 were excluded due to severe artifacts (n=7, 7.3%), different CT contrast injection protocols (n=11, 11.5%), heart failure (n=5, 5.2%), or technical failure of the contrast injection (n=8, 8.3%). Thus, the final cohort comprised 65 patients (41 males, 24 females; mean age: 59±16 years; range, 33–78 years). Each patient underwent two consecutive CT scans during a single-breath hold in the portal venous phase in accordance with a standard-dose protocol (group A) and an ULD protocol (group B) (5). The standard-dose protocol aimed to achieve a volume computed tomography dose index (CTDI_vol) between the 25^th to 75^th percentile, while the ULD protocol targeted the 25^th percentile CTDI_vol, of the suggested radiation dose range of 9 to 19 mGy by the American College of Radiology Dose Index Registry (17).

Image acquisition and processing

A 512-slice spiral CT instrument (NeuViz Epoch, Neusoft Medical Systems Co., Ltd., Shenyang, China) was used for scanning. The standard-dose and ULD protocols were performed sequentially, first in the craniocaudal direction, then in the reverse direction in the same breath hold during the portal venous phase (5). The duration of the venous phase scan was around 3 s, and the time interval between two consecutive scans during the portal venous phase was around 1.7 s. This design was intended to maintain stable respiratory conditions and minimize inter-scan variability. The sequence of scans was alternated among participants. The standard-dose protocol used a standard tube voltage of 120 kV with automatic tube current modulation (O-Dose) and a predefined signal-to-noise ratio (SNR) of 1.0 during the portal venous phase. The SNR parameter was set to determine a protocol-specific reference water-equivalent diameter (e.g., 330 mm for abdominal imaging) and reference milliampere seconds (mAs), e.g., 200 mAs, ensuring consistent image quality across anatomically diverse patients. This parameter (a default SNR of 1.0 for 330 mm) served as the primary determinant for radiation dose modulation, allowing protocol-specific manual adjustments where higher SNR values prioritize image quality at the expense of an increased dose. Conversely, the ULD protocol employed a reduced SNR of 0.5, resulting in a proportional reduction in the radiation dose. The other imaging parameters included a spiral pitch of 0.8, rotation time of 0.5 s/r, slice thickness and interval of 1.0 mm, reconstruction matrix of 512×512, collimation width of 256×0.625 mm, and filter parameter of F20.

The non-ionic contrast agent (Iohexol, Omnipaque 350 mgI/mL, GE Healthcare, WI, USA) was injected using a power injector (Envision CT injector, Medrad, Indianola, PA, USA) via a 22-G catheter inserted into the median cubital vein. A weight-based contrast dose protocol of 1.2 mL/kg was used, with the contrast dose ranging from 50 to 95 mL, and an injection speed of 3.0 mL/s. A bolus tracking technique was used to monitor the area of interest, with a trigger threshold set at 100 Hounsfield units (HU) at the level of the abdominal aorta. The arterial phase commenced 12 s after triggering, followed by the portal venous phase after a 30-s delay.

Principle of the deep-learning algorithm and CT image reconstruction

The advanced DLIR technique used in our study, which is a commercially available vendor-specific DLIR algorithm, employs a restructuring engine based on deep neural networks. This algorithm processes raw projection data from CT scanners, capturing internal anatomical structures at various angles. To mitigate the effect of detector noise, the algorithm incorporates noise characteristics into its processing. Additionally, it considers the CT physical model, including factors such as scattering, beam hardening, focal spot size, and detector dimensions, which influence image quality. The forward projection step emulates the natural X-ray imaging process. The optimization objective, comprising a data consistency term and a regularization term facilitated by a deep-learning model, ensures the reconstructed image aligns with the original data and enhances image quality by incorporating prior knowledge. The deep-learning model operates in the image domain, extracting features to refine details and reduce noise, resulting in high-quality CT images with enhanced clarity and contrast. Back projection iteratively refines the image to match the original data, achieving superior reconstruction. A schematic of the network structure and algorithm is shown in Figure 2. For image reconstruction, 10 different weight levels (ranging from 0% to 90% in 10% intervals) for DLIR are available for users to select, where higher percentages indicate stronger noise reduction from the deep-learning algorithm. In this study, a medium level of 50% was selected to achieve a moderate noise reduction based on the manufacturer’s recommendations and previous preliminary experiments (18).

Figure 2 Illustrative diagram of the ClearInfinity algorithm based on DLIR. (A) The network structure diagram of DLIR. (B) The schematic diagram of DLIR. CT, computed tomography; DLIR, deep-learning image reconstruction.

In group A (standard dose), the images were reconstructed with a 1.0-mm thickness using ClearView (CV, Neusoft medical, Shenyang, China) at a 50% weight, which represents a standard IR protocol. In group B (ULD), the images were reconstructed using 50% CV for group B1, and using the DLIR algorithm (CI) at a 50% weight for group B2. The reconstruction speed for each algorithm was also recorded, starting from the display of the first image and ending with the last image. The total time taken (T) and the total number of images (N) were recorded, and the image reconstruction speed was calculated as (N − 1)/T, with the unit of measurement in images per second (ips).

Quantitative image analysis

A technician with 10 years of experience in abdominal CT, who was blinded to the image review results, performed the quantitative measurements at a commercially available workstation (AVW 2.0, Neusoft medical, Shenyang, China). Following an existing procedure (19) for attenuation measurements, the regions of interest (ROIs) were placed within the hepatic parenchyma, pancreas, portal vein, and paraspinal muscle at the level of the main portal vein. Liver attenuation was recorded as the mean measurement value of three ROIs in the right anterior, right posterior, and left section of the liver, respectively. Portal vein attenuation was recorded based on a single, hand-drawn ROI placed at the right and left portal vein confluence level. For pancreatic attenuation, the CT value was obtained using a single ROI drawn at a spared portion of the pancreatic parenchyma, while carefully avoiding the main pancreatic duct, visible vessels, and artifacts. The attenuation of the paraspinal muscle was recorded, while also avoiding macroscopic fat infiltration. Each CT value was calculated by averaging the three-time measurements. Image noise was defined as the standard deviation (SD) of the HU in a homogeneous region of the subcutaneous fat on the anterior abdominal wall, away from artifacts or vessels. Illustrations of ROI selection are presented in Figure S1.

For each set of images, the contrast-to-noise ratios (CNRs) for the ROIs were calculated using the following formula:

CNRROI=CTROI−CTmSDn[1]

where CT_ROI and CT_m are the mean attenuations of the organ of interest and the paraspinal muscle, respectively, and SD_n is the mean image noise, defined as the SD value of the subcutaneous fat in the abdominal wall.

Qualitative image analysis

Two radiologists, each with more than five years of work experience, independently evaluated images from the three groups (A, B1, and B2) using a double-blind approach. The images were randomly displayed in a soft-tissue window with a window width of 400 HU and a window level of 40 HU to evaluate image quality. Each radiologist adjusted the window width and level based to their experience. Image noise, artifacts, and overall image quality were scored using a 5-point Likert scale on which 5 represented excellent, 4 represented good, 3 represented acceptable, 2 represented sub-acceptable, and 1 represented not acceptable (20,21). The evaluation criteria are listed in Appendix 1. Prior to assessment, the readers were informed that an image quality score of less than or equal to 2 would be deemed inadequate for the diagnosis.

Lesion detection and reference standards

After assessing the image quality metrics, the same radiologists independently reviewed all images on the same diagnostic workstation in a double-blind manner. Lesion visualization was subjectively evaluated using a 5-point ordinal scale (22) on which 1 indicated severely reduced image quality, rendering reliable interpretation impossible; 2 indicated severe blurring and poorly defined structures, resulting in uncertainty in the evaluation; 3 indicated moderate blurring of the interface structures, resulting in a slightly restricted assessment; 4 indicated slight blurring of the structures, allowing for unrestricted image evaluation; and 5 indicated excellent image quality with clear demarcation of structures.

All the datasets were displayed using a soft-tissue setting (width/level, 400/40 HU), with the option to adjust the window settings, zoom, and pan during review. In cases of disagreement during image analysis, a consensus reading was conducted. The consensus was determined by two other non-blinded, board-certified radiologists (with 15 and 19 years of abdominal imaging experience, respectively). The number and size of visible standard lesions were recorded using electronic calipers. The lesions were further categorized into low- and/or high-contrast lesions based on the CT attenuation relative to the surrounding tissues (for details, see Appendix 2) (23,24).

The reference standard was established by the two consensus reviewers using all the available clinical data and saved reader marks. If a disagreement arose between the consensus reviewers, the final decision was made by a senior abdominal radiologist (with 38 years of experience). Lesion diagnoses were established through a combination of all available cross-sectional imaging examinations (such as CT, magnetic resonance imaging, or positron emission tomography/CT) with at least 6 months of follow-up and/or with pathological examinations (biopsy or surgery). A lesion was considered detectable if identified by either reader on a given image. In terms of the performance metrics, lesions identified by the reference standard but not detected by the reviewers were classified as “false negatives”. Conversely, lesions that were marked but did not actually exist, such as pseudo-lesions in the falciform ligament, low-density artifacts caused by heterogeneous fatty liver, or image noise, were considered “false positives”. The lesion detection rate was calculated using the following formula:

lesiondetectionrate=NvisiblelesionsNtotallesions×100%[2]

Radiation dose

The CTDI_vol and dose-length product (DLP) were automatically calculated and recorded during CT scanning for both groups A and B. The ED was derived from the DLP using a conversion factor of 0.015 for abdominal examination, as recommended by the International Commission on Radiological Protection publication 103 (25). Further, the size-specific dose estimate (SSDE) was calculated according to the methods outlined in American Association of Physicists in Medicine (AAPM) Report 204, which involves the multiplication of the CTDI_vol by a size-specific conversion factor (26).

Statistical analysis

The SPSS statistical software (Windows v.23.0; SPSS, Chicago, Illinois) was used for the statistical analyses. A significance level of P<0.05 was used to determine statistical significance. The normality of the data distribution for objective image quality parameters (CT value, image noise, SNR, and CNR) and radiation dose (CTDI_vol, DLP, ED, and SSDE) was assessed using the Shapiro-Wilk test. The normally distributed data are presented as the mean ± SD, while the skewed data are expressed as the median (upper and lower quartiles). The categorical variables are expressed as the frequency. A repeated-measures analysis of variance was conducted to evaluate the quantitative image quality parameters, including CT values, image noise, the SNR, and the CNR. The qualitative image scores were compared using Friedman’s test. In cases of significant differences, post-hoc pairwise comparisons with Bonferroni correction were applied. The lesion detection rate was analyzed using Cochran’s Q test, with subsequent post-hoc pairwise McNemar tests conducted for significant outcomes. The sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated based on the confusion matrix. Inter-reader agreement was assessed using Cohen’s kappa (κ) statistics, with interpretations of κ values categorized as follows: 0.81–1.00 (excellent), 0.61–0.80 (substantial), 0.41–0.60 (moderate), 0.21–0.40 (fair), and 0.00–0.20 (poor). All the statistical analyses were conducted using industry-standard statistical software, adhering to best practices in statistical reporting for scientific rigor.

Results

Participant population

The study included 65 patients (41 males and 24 females) with a mean age of 59±16 years (range, 33–78 years) and a mean weight of 65±10 kg (range, 46–90 kg). The CT scans identified the following diseases among the 65 patients: hepatic metastases (n=15), hepatic cysts (n=8), hepatic carcinomas (n=13), cirrhoses (n=10), hepatic hemangiomas (n=8), cholangiocarcinomas (n=2), and gallstones (n=6); in three cases, no focal hepatic lesions were identified (n=3). The demographic characteristics of the patients are summarized in detail in Table 1.

Quantitative image analysis

In terms of the quantitative comparison (Table 2), the CT values and CNRs were similar in groups B2 and A (all P>0.05). However, the mean CNR values were significantly higher in group B2, increasing by 29.9–42.2%, than group B1 (all P<0.001). In terms of the image noise, group B1 had the highest mean noise levels at 10.05±2.94 HU, while group A had levels of 8.29±2.82 HU, and group B2 had levels of 8.04±2.71 HU. There was no significant difference between groups A and B2 in terms of image noise (P=0.625; Figure 3).

Figure 3 Box plots of image noise among the three image groups. The central line represents the median, the box boundaries indicate the first and third quartiles, and the whisker boundaries extend to 1.5 quartiles. Other points outside the whisker boundaries are plotted as outliers. Group A: standard-dose protocol with 50% CV; group B1: ultra-low-dose protocol with 50% CV; group B2: ultra-low-dose protocol with 50% CI. CI, ClearInfinity; CV, ClearView; HU, Hounsfield units; SD, standard deviation.

Qualitative image analysis

The results of the qualitative image quality evaluation are summarized in Table 3. The inter-rater agreement for evaluations of image noise, artifacts, and overall image quality was satisfactory (all κ coefficients >0.7). There were no significant differences between group A and group B2 across various image quality metrics (P>0.05). All the images in groups A and B2 were deemed diagnostically acceptable (scores ≥3). Conversely, all the images in group B1 were deemed unacceptable (all scores <3).

Lesion detection

Using the reference standard, 147 lesions (25 hepatic cysts, 8 hepatic hemangiomas, 19 hepatocellular carcinomas, 84 hepatic metastases, 7 cholangiocarcinomas, and 4 gallstones) in the enrolled patients were evaluated.

For the hepatic lesions, 144 of 147 lesions (98.0%) were detected in group A. While the detection rates for groups B1 and B2 were 81.6% (120/147) and 90.5% (133/147), respectively (P<0.001). For the high-contrast hepatic lesions, both detection rates and sensitivities reached 100% across all protocols with no statistically significant differences (P>0.05). Conversely, the performance for the low-contrast lesions varied significantly depending on the protocol. The lesion detection rates, sensitivities, specificities, PPVs, and NPVs for the low-contrast lesions among the three groups were evaluated based on the lesion type and size (as detailed in Table 4 and Figure 4). The results revealed that group A achieved the highest diagnostic accuracy (sensitivity: 97.2%; specificity: 85.4%; PPV: 88.3%; NPV: 96.5%), followed by group B2, which showed intermediate performance (87.2%, 81.3%, 84.1%, and 84.8%, respectively), and group B1, which had the lowest accuracy. The lower detection rate in the ULD protocol was primarily due to the lower detectability of small lesions (<0.5 cm, 31/54, 57.4% in group B1, and 42/54, 77.8% in group B2) and low-contrast lesions (82/109, 75.2% in group B1, and 95/109, 87.2% in group B2). Notably, group B2 had significantly higher detection rates for small or low-contrast lesions than group B1 (P<0.05), while no significant differences were observed in relation to the medium or large hepatic lesions (all P>0.05). Figure 5 illustrates a representative case.

Figure 4 Comparison of lesion detection rates among the three groups. (A) Comparison of lesion detection rates by lesion type. (B) Comparison of lesion detection rates by lesion size. *, P<0.05; **, P<0.001. Group A: standard-dose protocol with 50% CV; group B1: ultra-low-dose protocol with 50% CV; group B2: ultra-low-dose protocol with 50% CI. CI, ClearInfinity; CV, ClearView.

Figure 5 A 42-year-old female with hepatic metastases from cervical cancer was evaluated using the following three protocols: (A) SDCT with 50% CV reconstruction; (B) ULD-CT with 50% CV reconstruction; (C) ULD-CT with 50% CI reconstruction. An arrow indicates the hepatic metastatic lesion in each image. The image noise levels for the three groups were 9.17, 12.66, and 9.38 HU, respectively. Compared to the SDCT image (A), the ULD-CT image with 50% CI reconstruction (C) showed a 76.1% reduction in the radiation dose, while preserving the lesion’s structure, details and clear boundaries to meet the clinical diagnostic requirements, while the ULD-CT with 50% CV (B) had higher image noise and less clear lesion boundaries. CI, ClearInfinity; CT, computed tomography; CV, ClearView; HU, Hounsfield units; SDCT, standard-dose CT; ULD-CT, ultra-low-dose CT.

The inter-rater agreement for lesion visualization evaluations was moderate (k value range, 0.69–0.82). There was a significant difference in the lesion visualization score among the three groups (refer to Table 3), but no significant difference was found between groups A and B2 (P=0.189).

Radiation dose and reconstruction time

For the standard-dose protocol (group A), the mean CTDI_vol was 11.6±5.8 mGy (range, 3.8–26.3 mGy), the mean SSDE was 16.94±8.53 mGy (range, 5.25–34.7 mGy), and the mean ED was 6.9±2.0 mSv (range, 1.68–13.03 mSv). In the ULD protocol (group B), the mean CTDI_vol was 3.1±2.0 mGy (range, 1.2–7.7 mGy), the mean SSDE was 4.51±2.82 mGy (range, 1.28–8.48 mGy), and the mean ED was 1.5±0.8 mSv (range, 0.57–4.14 mSv). The mean CTDI_vol, SSDE, and ED differed significantly between the two groups (P<0.001), and this corresponded to a 73.3% reduction in the estimated dose. The multi-detector computed tomography (MDCT) scanner processed data using the DLIR engine in a single iteration to generate CI images. The DLIR engine achieved an average reconstruction rate of 60 ips, which closely matched the performance of the CV system, which also operated at 60 ips.

Discussion

Our study evaluated the image quality and lesion detection capabilities of ULD contrast-enhanced abdominal CT using a novel real-time DLIR technique (CI) in patients with hepatic lesions. The results revealed a remarkable 73.3% reduction in the radiation dose with the 50% CI algorithm in the ULD-CT, while maintaining comparable image quality, CNRs, and lesion detectability for liver low-contrast lesions larger than 0.5 cm, compared to the standard-dose protocol with the 50% CV algorithm. Moreover, the images produced by the 50% CI algorithm exhibited improved image quality and sensitivity in detecting small-size or low-contrast lesions compared to those reconstructed with the 50% CV algorithm in the ULD protocol.

Image noise is a critical factor influencing diagnostic accuracy, especially in identifying small or low-contrast lesions in enhanced CT images, where heightened noise poses a greater challenge. IR algorithms, when applied at high strengths, may produce excessively smooth images, potentially impairing the detection of low-contrast lesions and thus compromising diagnostic efficacy (6,7). In our study, group B1 (ULD protocol with 50% CV) had high image noise and artifacts, resulting in blurred organ boundaries and compromised diagnostic confidence. These findings align with those of previous studies (27,28), emphasizing the need for careful consideration when applying the IR algorithm in clinical practice.

DLIR technology (29,30) has revolutionized noise reduction and structural preservation in low-dose CT; however, detecting hepatic lesions with a reduced radiation dose remains a challenge due to the compromised image contrast (13,31). Our study introduced the novel CI algorithm, and showed its efficacy in ULD protocols for hepatic lesion detection. The CI algorithm achieved a lesion detection rate of 90.5%, with higher diagnostic sensitivity across various lesion sizes and types. These results are consistent with those of previous studies. For example, Noda et al. (15) highlighted the effectiveness of DLIR in low-dose CT primarily for high-contrast lesions. Notably, our study expanded on these findings by including both high-contrast and low-contrast lesions, further validating the versatility of the CI algorithm. Lesion visualization scores for the 50% CI images in the ULD protocol were significantly higher than those for the 50% CV images (4.18±0.90 vs. 2.86±0.61, P<0.001) and comparable to those obtained with the standard-dose 50% CV images (4.38±0.58, P=0.189). Additionally, the CI algorithm demonstrated superior capability in detecting low-contrast hepatic lesions larger than 0.5 cm, and outperformed conventional CV algorithms in identifying lesions smaller than 0.5 cm. These findings suggest that the novel DLIR algorithm not only enhances image quality but also maintains diagnostic accuracy in ULD-CT protocols. This makes it particularly valuable for clinical applications requiring frequent imaging, such as in oncology follow up.

Beyond its diagnostic performance, the CI algorithm exhibited significantly reduced reconstruction times compared to other DLIR approaches (32-34). Specifically, the reconstruction time in our study was 60 ips, which was comparable to that of the IR algorithm, while the reconstruction times for other DLIR algorithms were increased by 17 s compared to the IR algorithm (32). This improvement is attributed to advanced optimization strategies, including enhancements in model architecture, parameter tuning, and computational efficiency. These refinements minimized the system response latency during the algorithm inference, enabling faster image reconstruction. Such efficiency is especially advantageous in high-throughput clinical environments, such as large-volume hospitals, where workflow optimization is critical. In summary, the implementation of DLIR-assisted ULD-CT with the CI algorithm resulted in a substantial 73.3% reduction in the radiation dose, while preserving image quality and diagnostic accuracy. This approach effectively addresses the dual imperatives of minimizing patient radiation exposure and enhancing operational efficiency. It represents a significant advancement in routine imaging protocols, particularly for oncology patients requiring regular follow-up imaging.

This study had several limitations. First, the sample size was small, necessitating further validation to confirm the results. Second, all patients enrolled were scanned using 120 kV. However, for patients with a body mass index less than 25 kg/m², a lower tube voltage of 100 kV or less might be more appropriate. Third, although the standard-dose and ULD scans were acquired sequentially in a single-breath hold (with a 1.7-s inter-scan interval), non-simultaneous acquisition could still introduce subtle anatomical displacement. Finally, our comparison focused exclusively on the image quality between the 50% CV and 50% CI algorithms based on previous phantom experiments and manufacturer recommendations. Future studies should consider assessing other algorithms and varying weights to provide a more comprehensive understanding of the topic.

Conclusions

Compared to standard-dose abdominal CT, the application of the novel real-time DLIR technique resulted in a significant 73.3% reduction in the radiation dose. This approach effectively minimizes image noise while maintaining clinically acceptable image quality, ensuring diagnostic confidence, particularly for the detection of hepatic lesions larger than 0.5 cm.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STROBE reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-365/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-365/dss

Funding: This study was supported by the National Natural Science Foundation of China (grant No. 72271060), and the Young Elite Scientists Sponsorship Program by CAST (grant No. 2022QNRC001).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-365/coif). C.X. is an employee of the Neusoft Medical System Company, the manufacturer of the CT system used in this study. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Human Research Ethics Committee of The First Affiliated Hospital of Zhengzhou University (No. 2022-KY-0752-001). All patients provided written informed consent before participation.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Pickhardt PJ, Graffy PM, Said A, Jones D, Welsh B, Zea R, Lubner MG. Multiparametric CT for Noninvasive Staging of Hepatitis C Virus-Related Liver Fibrosis: Correlation With the Histopathologic Fibrosis Score. AJR Am J Roentgenol 2019;212:547-53. [Crossref] [PubMed]
2. Wulff AM, Bolte H, Fischer S, Freitag-Wolf S, Soza G, Tietjen C, Biederer J, Heller M, Fabel M. Lung, liver and lymph node metastases in follow-up MSCT: comprehensive volumetric assessment of lesion size changes. Rofo 2012;184:820-8. [Crossref] [PubMed]
3. Tezcan D, Türkvatan A, Türkoğlu MA, Bostancı EB, Sakaoğullları Z. Preoperative staging of colorectal cancer: accuracy of single portal venous phase multidetector computed tomography. Clin Imaging 2013;37:1048-53. [Crossref] [PubMed]
4. Baskan O, Erol C, Ozbek H, Paksoy Y. Effect of radiation dose reduction on image quality in adult head CT with noise-suppressing reconstruction system with a 256 slice MDCT. J Appl Clin Med Phys 2015;16:5360. [Crossref] [PubMed]
5. Jensen CT, Gupta S, Saleh MM, Liu X, Wong VK, Salem U, Qiao W, Samei E, Wagner-Bartak NA. Reduced-Dose Deep Learning Reconstruction for Abdominal CT of Liver Metastases. Radiology 2022;303:90-8. [Crossref] [PubMed]
6. McCollough CH, Yu L, Kofler JM, Leng S, Zhang Y, Li Z, Carter RE. Degradation of CT Low-Contrast Spatial Resolution Due to the Use of Iterative Reconstruction and Reduced Dose Levels. Radiology 2015;276:499-506. [Crossref] [PubMed]
7. Jensen K, Martinsen AC, Tingberg A, Aaløkken TM, Fosse E. Comparing five different iterative reconstruction algorithms for computed tomography in an ROC study. Eur Radiol 2014;24:2989-3002. [Crossref] [PubMed]
8. Benz DC, Benetos G, Rampidis G, von Felten E, Bakula A, Sustar A, Kudura K, Messerli M, Fuchs TA, Gebhard C, Pazhenkottil AP, Kaufmann PA, Buechel RR. Validation of deep-learning image reconstruction for coronary computed tomography angiography: Impact on noise, image quality and diagnostic accuracy. J Cardiovasc Comput Tomogr 2020;14:444-51. [Crossref] [PubMed]
9. Racine D, Becce F, Viry A, Monnin P, Thomsen B, Verdun FR, Rotzinger DC. Task-based characterization of a deep learning image reconstruction and comparison with filtered back-projection and a partial model-based iterative reconstruction in abdominal CT: A phantom study. Phys Med 2020;76:28-37. [Crossref] [PubMed]
10. Kim I, Kang H, Yoon HJ, Chung BM, Shin NY. Deep learning-based image reconstruction for brain CT: improved image quality compared with adaptive statistical iterative reconstruction-Veo (ASIR-V). Neuroradiology 2021;63:905-12. [Crossref] [PubMed]
11. Zheng Z, Ai Z, Liang Y, Li Y, Wu Z, Wu M, Han Q, Ma K, Xiang Z. Clinical value of deep learning image reconstruction on the diagnosis of pulmonary nodule for ultra-low-dose chest CT imaging. Clin Radiol 2024;79:628-36. [Crossref] [PubMed]
12. Hou P, Liu N, Feng X, Chen Y, Wang H, Wang X, Liu J, Zhan P, Liu X, Shang B, Shen Z, Wang L, Gao J, Lyu P. A paradigm shift in oncology imaging: a prospective cross-sectional study to assess low-dose deep learning image reconstruction versus standard-dose iterative reconstruction for comprehensive lesion detection in dual-energy computed tomography. Quant Imaging Med Surg 2024;14:6449-65. [Crossref] [PubMed]
13. Lyu P, Liu N, Harrawood B, Solomon J, Wang H, Chen Y, Rigiroli F, Ding Y, Schwartz FR, Jiang H, Lowry C, Wang L, Samei E, Gao J, Marin D. Is it possible to use low-dose deep learning reconstruction for the detection of liver metastases on CT routinely? Eur Radiol 2023;33:1629-40. [Crossref] [PubMed]
14. Cao L, Liu X, Li J, Qu T, Chen L, Cheng Y, Hu J, Sun J, Guo J. A study of using a deep learning image reconstruction to improve the image quality of extremely low-dose contrast-enhanced abdominal CT for patients with hepatic lesions. Br J Radiol 2021;94:20201086. [Crossref] [PubMed]
15. Noda Y, Kaga T, Kawai N, Miyoshi T, Kawada H, Hyodo F, Kambadakone A, Matsuo M. Low-dose whole-body CT using deep learning image reconstruction: image quality and lesion detection. Br J Radiol 2021;94:20201329. [Crossref] [PubMed]
16. Mileto A, Zamora DA, Alessio AM, Pereira C, Liu J, Bhargava P, et al. CT Detectability of Small Low-Contrast Hypoattenuating Focal Lesions: Iterative Reconstructions versus Filtered Back Projection. Radiology 2018;289:443-54. [Crossref] [PubMed]
17. Kanal KM, Butler PF, Sengupta D, Bhargavan-Chatfield M, Coombs LP, Morin RL U.S.. Diagnostic Reference Levels and Achievable Doses for 10 Adult CT Examinations. Radiology 2017;284:120-33. [Crossref] [PubMed]
18. Qi K, Xu C, Yuan D, Zhang Y, Zhang M, Zhang W, Zhang J, You B, Gao J, Liu J. Feasibility of Ultra-low Radiation and Contrast Medium Dosage in Aortic CTA Using Deep Learning Reconstruction at 60 kVp: An Image Quality Assessment. Acad Radiol 2025;32:1506-16. [Crossref] [PubMed]
19. Marin D, Nelson RC, Schindera ST, Richard S, Youngblood RS, Yoshizumi TT, Samei E. Low-tube-voltage, high-tube-current multidetector abdominal CT: improved image quality and decreased radiation dose with adaptive statistical iterative reconstruction algorithm--initial clinical experience. Radiology 2010;254:145-53. [Crossref] [PubMed]
20. Nakaura T, Nakamura S, Maruyama N, Funama Y, Awai K, Harada K, Uemura S, Yamashita Y. Low contrast agent and radiation dose protocol for hepatic dynamic CT of thin adults at 256-detector row CT: effect of low tube voltage and hybrid iterative reconstruction algorithm on image quality. Radiology 2012;264:445-54. [Crossref] [PubMed]
21. Wang X, Zheng F, Xiao R, Liu Z, Li Y, Li J, Zhang X, Hao X, Zhang X, Guo J, Zhang Y, Xue H, Jin Z. Comparison of image quality and lesion diagnosis in abdominopelvic unenhanced CT between reduced-dose CT using deep learning post-processing and standard-dose CT using iterative reconstruction: A prospective study. Eur J Radiol 2021;139:109735. [Crossref] [PubMed]
22. Zeng L, Xu X, Zeng W, Peng W, Zhang J, Sixian H, Liu K, Xia C, Li Z. Deep learning trained algorithm maintains the quality of half-dose contrast-enhanced liver computed tomography images: Comparison with hybrid iterative reconstruction: Study for the application of deep learning noise reduction technology in low dose. Eur J Radiol 2021;135:109487. [Crossref] [PubMed]
23. Solomon J, Marin D, Roy Choudhury K, Patel B, Samei E. Effect of Radiation Dose Reduction and Reconstruction Algorithm on Image Noise, Contrast, Resolution, and Detectability of Subtle Hypoattenuating Liver Lesions at Multidetector CT: Filtered Back Projection versus a Commercial Model-based Iterative Reconstruction Algorithm. Radiology 2017;284:777-87. [Crossref] [PubMed]
24. Inoue A, Voss BA, Lee NJ, Takahashi H, Kozaka K, Heiken JP, Ehman EC, Vasconcelos R, Fidler JL, Lee YS, Mileto A, Johnson MP, Baer-Beck M, Weber NM, Michalak GJ, Halaweish A, Carter RE, McCollough CH, Fletcher JG. Diagnostic Performance in Low- and High-Contrast Tasks of an Image-Based Denoising Algorithm Applied to Radiation Dose-Reduced Multiphase Abdominal CT Examinations. AJR Am J Roentgenol 2023;220:73-85. [Crossref] [PubMed]
25. Christner JA, Kofler JM, McCollough CH. Estimating effective dose for CT using dose-length product compared with using organ doses: consequences of adopting International Commission on Radiological Protection publication 103 or dual-energy scanning. AJR Am J Roentgenol 2010;194:881-9. [Crossref] [PubMed]
26. Li B, Behrman RH. Comment on the "report of AAPM TG 204: size-specific dose estimates (SSDE) in pediatric and adult body CT examinations" Med Phys 2012;39:4613-4; author reply 4615-6. [report of AAPM TG 204, 2011].
27. Singh R, Digumarthy SR, Muse VV, Kambadakone AR, Blake MA, Tabari A, Hoi Y, Akino N, Angel E, Madan R, Kalra MK. Image Quality and Lesion Detection on Deep Learning Reconstruction and Iterative Reconstruction of Submillisievert Chest and Abdominal CT. AJR Am J Roentgenol 2020;214:566-73. [Crossref] [PubMed]
28. Park J, Shin J, Min IK, Bae H, Kim YE, Chung YE. Image Quality and Lesion Detectability of Lower-Dose Abdominopelvic CT Obtained Using Deep Learning Image Reconstruction. Korean J Radiol 2022;23:402-12. [Crossref] [PubMed]
29. Chen H, Zhang Y, Kalra MK, Lin F, Chen Y, Liao P, Zhou J, Wang G, Low-Dose CT. With a Residual Encoder-Decoder Convolutional Neural Network. IEEE Trans Med Imaging 2017;36:2524-35. [Crossref] [PubMed]
30. Kang E, Chang W, Yoo J, Ye JC. Deep Convolutional Framelet Denosing for Low-Dose CT via Wavelet Residual Network. IEEE Trans Med Imaging 2018;37:1358-69. [Crossref] [PubMed]
31. Nakamura Y, Higaki T, Tatsugami F, Zhou J, Yu Z, Akino N, Ito Y, Iida M, Awai K. Deep Learning-based CT Image Reconstruction: Initial Evaluation Targeting Hypovascular Hepatic Metastases. Radiol Artif Intell 2019;1:e180011. [Crossref] [PubMed]
32. Oostveen LJ, Meijer FJA, de Lange F, Smit EJ, Pegge SA, Steens SCA, van Amerongen MJ, Prokop M, Sechopoulos I. Deep learning-based reconstruction may improve non-contrast cerebral CT imaging compared to other current reconstruction algorithms. Eur Radiol 2021;31:5498-506. [Crossref] [PubMed]
33. Tamura A, Mukaida E, Ota Y, Kamata M, Abe S, Yoshioka K. Superior objective and subjective image quality of deep learning reconstruction for low-dose abdominal CT imaging in comparison with model-based iterative reconstruction and filtered back projection. Br J Radiol 2021;94:20201357. [Crossref] [PubMed]
34. Koetzier LR, Mastrodicasa D, Szczykutowicz TP, van der Werf NR, Wang AS, Sandfort V, van der Molen AJ, Fleischmann D, Willemink MJ. Deep Learning Image Reconstruction for CT: Technical Principles and Clinical Prospects. Radiology 2023;306:e221257. [Crossref] [PubMed]