Comparison of image quality in 40 keV virtual monoenergetic images of dual-energy CT pulmonary angiography using deep learning and iterative reconstruction algorithms under optimized low dose scanning protocols

Introduction

Pulmonary embolism is a prevalent and potentially life-threatening cardiovascular condition, characterized by the obstruction of pulmonary arteries due to embolic material (1). Early and accurate diagnosis is critical, as timely intervention can significantly reduce morbidity and mortality (2,3). Computed tomography pulmonary angiography (CTPA) is widely regarded as the gold standard for diagnosing pulmonary embolism, owing to its ability to quickly provide high-resolution images of the pulmonary vasculature, thereby enabling the detection of even small emboli with high sensitivity and specificity (4).

However, traditional single energy CT angiography, while effective, necessitates the use of high radiation doses and large volumes of iodine contrast agents to optimize vascular contrast and signal-to-noise ratio (SNR). These increases in radiation and contrast agent volumes, while improving diagnostic capability, concurrently elevate the risks of contrast-induced nephropathy (CIN) and radiation-induced tissue damage. In addition, the rapid injection of high-concentration contrast agents can increase the likelihood of extravasation, particularly in patients with vascular pathology such as arterial sclerosis and post-chemotherapy. Thus, although traditional CTPA are highly effective, they come with inherent risks that necessitate the exploration of alternative imaging strategies.

In response to these concerns, dual-energy CTPA (DE-CTPA) has emerged as a promising alternative. DE-CTPA utilizes two different X-ray energy levels to generate virtual monoenergetic images (VMI), which may enhance the effectiveness of iodine contrast while reducing the required contrast volume (5,6). This modality not only can reduce the amount of contrast agent but also increase the contrast of the blood vessels. Nevertheless, while DE-CTPA has clear advantages, one challenge associated with low-kilo electron volt (keV) imaging (such as at 40 keV) is the significant increase in image noise, which diminishes diagnostic accuracy and impedes the detection of small embolic lesions or subtle vascular abnormalities.

In dual-energy CT, low-keV VMI can enhance iodine attenuation because of their proximity to the iodine K-edge at 33.2 keV. Among these, 40 keV VMI provide the strongest vascular contrast, which is particularly advantageous under reduced contrast medium protocols. However, this advantage is counterbalanced by a substantial increase in image noise, often limiting their routine clinical use. Higher monoenergetic levels (50–70 keV) can effectively mitigate image noise, but they also markedly decrease iodine conspicuity, potentially reducing the visibility of small intravascular filling defects. Phantom task-based evaluations have similarly shown that while noise decreases with increasing keV, lesion detectability peaks around 40–50 keV, particularly when deep learning reconstruction is applied (7). In this study, we therefore focused on 40 keV VMI as a representative “high-contrast but high-noise” condition, hypothesizing that deep learning image reconstruction (DLIR) could counteract the noise penalty and render such images diagnostically reliable.

To address this issue, iterative reconstruction algorithms like Adaptive Statistical Iterative Reconstruction-V (ASIR-V) have been introduced. These algorithms utilize advanced noise modeling techniques to optimize image quality, particularly under low-dose conditions. ASIR-V is effective in reducing noise, enhancing clarity, and improving the overall diagnostic quality of the images. However, one limitation of ASIR-V is its tendency to reduce the natural texture of images, which may result in a loss of fine anatomical details, particularly in the vascular structures that are critical for accurate interpretation (8).

A more recent advancement, DLIR algorithms, has gained attention for its potential to overcome some of these challenges (9). Unlike traditional iterative methods, DLIR, based on convolutional neural networks (CNNs), has been shown to provide superior noise reduction while preserving fine anatomical details. This approach offers a promising solution for low-dose CT imaging, particularly in preserving image quality at lower radiation doses (10,11). Despite its potential, the application of DLIR in dual-energy, low-keV VMI imaging for low-dose DE-CTPA remains underexplored, with limited empirical evidence supporting its effectiveness in this specific clinical context.

Thus, the primary aim of this study is to compare the image quality of 40 keV VMI images reconstructed using DLIR and ASIR-V algorithms under low-dose DE-CTPA protocols.

Methods

General information

This prospective study enrolled patients who underwent DE-CTPA at our institution between January and April 2025. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of the Affiliated Hospital of Xuzhou Medical University (approval No. XYFY2025-KL572-01) and informed consent was obtained from all individual participants.

Inclusion criteria

(I) Clinical suspicion of pulmonary embolism, necessitating CTPA for diagnostic confirmation; (II) a confirmed diagnosis of pulmonary embolism requiring thrombolytic therapy, followed by follow-up imaging for therapeutic evaluation.

Exclusion criteria

(I) Known allergies to iodine-based contrast agents; (II) severe cardiac, hepatic, or renal dysfunction that contraindicated the use of contrast agents or compromised image quality; (III) hyperthyroidism; (IV) inability to cooperate with the examination due to significant agitation or other factors (Figure 1).

Figure 1 Flowchart of the enrolled patients in this prospective study. ASIR-V, Adaptive Statistical Iterative Reconstruction-V; CNR, contrast-to-noise ratio; CT, computed tomography; CTPA, computed tomography pulmonary angiography; DLIR-H, deep learning image reconstruction with high setting; DLIR-M, deep learning image reconstruction with medium setting; GSI, Gemstone Spectral Imaging mode; HU, Hounsfield unit; SD, standard deviation; SNR, signal-to-noise ratio.

Clinical data

The clinical data, including demographic information (age, sex), body mass index (BMI), contrast agent volume, and injection rate, were collected by a radiologist (L.Z.) with 2 years of experience in cardiovascular imaging. Additionally, the CT reports documented CT volume dose index (CTDI_vol) and dose-length product (DLP), which are essential for estimating radiation dose exposure. The effective dose (ED) was subsequently calculated using the following equation:

ED=DLP×k[1]

where k=0.014 mSv/mGy·cm (12). This conversion factor was used to estimate the radiation dose to the patient, which is crucial for understanding the impact of imaging protocols on patient safety.

DE-CTPA scanning protocol

The DE-CTPA scans were conducted using a 256-row CT scanner (Revolution CT, GE Healthcare, Chicago 60661, USA). Patients were positioned in the supine position, and the scan was performed with a foot-to-head direction. The scan range extended from the thoracic inlet to the costophrenic angle to fully capture the pulmonary vasculature.

Iodixanol (350 mgI/mL) (Hengrui Pharmaceutical, Lianyungang 222000, China) was used as the contrast agent, administered via a high-pressure injector (CT Motion-XD8000, Erlangen, Germany) at a dose of 0.4 mL/kg over a 6-second period, followed by 30 mL of saline flush. Dynamic monitoring of the main pulmonary artery was carried out using smart tracking technology, with the scanning trigger set at 50 Hounsfield unit (HU) threshold. This technique ensured optimal imaging of the pulmonary arteries without overexposing the patient to unnecessary radiation.

The scanning was performed using Spectral CT mode (Gemstone Spectral Imaging mode, GSI mode) to enhance image quality while reducing the required contrast agent volume. The tube voltages were set at 80 and 140 kV, optimized for the tissue types encountered during DE-CTPA. The tube current was controlled by the GSI Assist program, with a reference current of 320 mA. The Noise index (NI), defined as the standard deviation of CT numbers in a uniform water phantom that determines the target image noise level and modulates the tube current accordingly, was set to 10, balancing high-quality imaging with minimal radiation exposure. NI is a key parameter for adjusting radiation dose. By controlling the NI value, the noise level in the image can be modulated. Higher NI values correspond to lower radiation dose, reducing patient exposure, while lower NI values improve image quality at the cost of higher radiation exposure. Adjusting the NI allows for an optimal balance between image quality and radiation dose to meet clinical needs (13).

The pitch was adjusted to 0.992:1, providing an optimal balance between scan time and image quality. The scan speed was 0.28 seconds per rotation, ensuring rapid data acquisition and minimizing motion artifacts. The detector width was 128×0.625 mm, selected for high-resolution imaging of fine vascular structures. Reconstruction thickness was set at 1.25 mm, with 0.625 mm intervals to ensure fine detail capture.

Image processing

The raw data were reconstructed into GSI data files using four distinct algorithms: ASIR-V 50% (Group 1), ASIR-V 70% (Group 2), deep learning image reconstruction with medium setting (DLIR-M) (Group 3), and deep learning image reconstruction with high setting (DLIR-H) (Group 4). And the GSI data files were transferred to the GE Advanced Workstation Server 4.7 (AW Server 4.7) for generating 40 keV VMI (Figures 2,3).

Figure 2 Comparison of image quality for 40 keV VMI reconstructed with different algorithms in a 66-year-old male patient (BMI, 31.23 kg/m²). Zoomed-in regions of the segmental pulmonary arteries demonstrate differences in noise level, vessel conspicuity, and image texture, with the algorithms directly indicated in the image. ASIR-V, Adaptive Statistical Iterative Reconstruction-V; BMI, body mass index; DLIR-H, deep learning image reconstruction with high setting; DLIR-M, deep learning image reconstruction with medium setting; VMI, virtual monoenergetic images.

Figure 3 Comparison of image quality for 40 keV VMI reconstructed with different algorithms in a 74-year-old female patient (BMI, 26.67 kg/m²) with pulmonary embolism. Low-density filling defects in the left and right pulmonary arteries, indicative of thrombus occlusion, are visible in both subjects. The differences in image quality between the reconstruction algorithms are highlighted, emphasizing their clinical relevance in pulmonary embolism diagnosis. ASIR-V, Adaptive Statistical Iterative Reconstruction-V; BMI, body mass index; DLIR-H, deep learning image reconstruction with high setting; DLIR-M, deep learning image reconstruction with medium setting; VMI, virtual monoenergetic images.

For ASIR-V, blending levels of 50% and 70% were selected as they are commonly used in clinical practice, offering a balance between effective noise reduction and preservation of natural image texture. Lower percentages (<40%) may not sufficiently reduce noise, while higher percentages (>80%) may result in overly smoothed images, giving them a waxy appearance.

The DLIR algorithm used in this study is a vendor-provided, pre-trained model developed by GE Healthcare (Milwaukee, WI, USA). It was trained on large-scale, multi-institutional CT image datasets encompassing diverse patient populations and imaging protocols, and does not require retraining on local datasets. Three selectable strength levels are available (low, medium, and high), which represent different balances between noise suppression and preservation of image texture. DLIR-M provides moderate noise reduction with preservation of spatial resolution, whereas DLIR-H applies stronger suppression, resulting in smoother image texture and higher SNR and CNR. The underlying network architecture and parameters are proprietary and not open source. The algorithm is designed to be applied directly in clinical practice without site-specific retraining (14,15).

The average reconstruction time was approximately 19 seconds per series for ASIR-V 50% and ASIR-V 70%, approximately 25 seconds for DLIR-M, and approximately 26 seconds for DLIR-H.

Image quality assessment

Objective image quality evaluation

A radiologist (J.L.) with 3 years of experience in cardiovascular imaging analyzed the reconstructed images using GSI Viewer software on the AW Server 4.7 workstation. The analysis of the arterial phase includes the CT attenuation and noise levels of the main pulmonary artery (PA), left pulmonary artery (LPA), right pulmonary artery (RPA), and subscapular muscles at the level of the pulmonary artery.

The region of interest (ROI) was placed centrally within the blood vessels, avoiding thrombus or artifacts. The ROI was chosen to cover approximately two-thirds of the vessel diameter, minimizing potential volume effects. The standard deviation (SD) of pixel values within the ROI served as the primary measure for image noise. Representative ROI placement images are provided in the supplementary file (Figure S1) to illustrate how measurements were performed in the main pulmonary artery, left pulmonary artery, right pulmonary artery, and subscapular muscle.

To assess image quality quantitatively, SNR and contrast-to-noise ratio (CNR) were calculated using the following formulas:

SNR=CTattenuationofpulmonaryarteryNoise[2]

CNR=CTattenuationofpulmonaryartery−CTattenuationofsubclavianmuscleNoise[3]

Subjective image quality evaluation

For subjective evaluation, a radiologist (J.L.) randomly selected images for independent assessment by two experienced radiologists (Y.M. and K.X.) with 15 and 30 years of expertise in CTA diagnostics, respectively. A double-blind method was employed to ensure objectivity in evaluation. The images were displayed on a JUSHA-C42E monitor (Nanjing Jusha Display Technology Co., Ltd., Nanjing, China) with a resolution of 2,560×1,440, and the evaluations took place in a controlled, distraction-free environment. Disagreements between the two radiologists were resolved through consensus.

Each image was evaluated using a 5-point Likert scale, assessing the following attributes: image noise, vascular edge clarity, vascular texture, and overall image quality (Table S1) (16). This scale allowed for both qualitative and quantitative assessment of image quality.

Statistical analysis

Statistical analyses were performed using SPSS 27.0 software. For normally distributed data, results are presented as the mean ± standard deviation. For non-normally distributed data, the median (interquartile range) is reported. Categorical data are expressed as percentages (%). One-way analysis of variance (ANOVA) was used for comparing normally distributed data with homogeneity of variance, with post-hoc pairwise comparisons conducted using the LSD test. For non-normally distributed data, the Kruskal-Wallis test was applied, followed by Dunn’s test for pairwise comparisons. Inter-rater reliability was assessed using intraclass correlation (ICC). An ICC ≥0.75 indicates excellent consistency, 0.60≤ ICC <0.75 suggests moderate consistency, 0.40≤ ICC <0.60 reflects fair consistency, and ICC <0.40 indicates poor consistency. Higher ICC values indicate greater reliability, essential for ensuring the validity of research findings. A P value of less than 0.05 was considered statistically significant.

Results

Patient demographics

A total of 75 patients were included in this study, comprising 39 males (52.0%) and 36 females (48.0%). The average age of the participants was 69.8±15.0 years, with a mean BMI of 24.24±2.89 kg/m². The cohort’s average ED was 3.76±1.02 mSv. The mean contrast usage volume administered was 26 mL, reflecting a 48% reduction compared to the conventional 50 mL dose typically used in standard CTPA protocols at our institution (Table 1).

Objective image quality evaluation

The image quality evaluation revealed significant differences between the four groups in terms of image noise, CNR, and SNR, which were measured at the PA, LPA, and RPA levels (P<0.001) (Table 2).

Post-hoc comparison of image quality between different algorithms

Post-hoc analysis revealed that, at the PA, LPA, and RPA levels, DLIR-M and DLIR-H exhibited significantly lower SD values than ASIR-V 50%, while CNR and SNR values were notably higher in DLIR-M and DLIR-H (P<0.001). Furthermore, DLIR-H displayed a significantly better performance compared to ASIR-V 70% with lower SD values and higher CNR and SNR (P<0.001). However, no significant differences were observed between ASIR-V 70% and DLIR-M in terms of SD value, CNR, or SNR (Table 3, Figure 4, Figure S2).

Figure 4 Post-hoc comparison of objective image quality evaluation (CNR, SNR). ASIR-V, Adaptive Statistical Iterative Reconstruction-V; CNR, contrast-to-noise ratio; DLIR-H, deep learning image reconstruction with high setting; DLIR-M, deep learning image reconstruction with medium setting; LA, left pulmonary artery; PA, pulmonary artery; RA, right pulmonary artery; SNR, signal-to-noise ratio.

Post-hoc comparison of image quality within same algorithms

Intra-algorithm comparisons revealed that within the ASIR-V group, ASIR-V 70% exhibited better image noise, CNR, and SNR compared to ASIR-V 50% (P<0.05). Within the DLIR algorithm, Group DLIR-H outperformed DLIR-M in terms of SD value, CNR, and SNR (P<0.01) (Table 3, Figure 4, Figure S2).

Subjective image quality evaluation

Interclass consistency for subjective image quality evaluation

To validate the subjective analysis, inter-rater consistency was assessed for image noise, spatial resolution, noise texture, and overall image quality. The ICC values for these variables were 0.887, 0.858, 0.852, and 0.855, respectively (Table S2).

Analysis of subjective image quality based on scoring

In terms of subjective ratings, DLIR-H achieved the highest proportion of scorning greater than 4 for image noise, spatial resolution, noise texture, and overall image quality (Figure S3).

Subjective image quality evaluation differences

Significant differences were observed across all four groups in image noise, spatial resolution, noise texture, and overall image quality (P<0.001) (Table 4).

DLIR-H consistently outperformed ASIR-V 50% and ASIR-V 70% in all categories (P<0.001). Additionally, DLIR-M showed better performance than ASIR-V 50% and ASIR-V 70% in spatial resolution and noise texture (P<0.05) (Table S3).

No significant statistical differences were observed between the subjective scoring of ASIR-V 50% and ASIR-V 70%.

Discussion

This study investigated the image quality of 40 keV VMI images reconstructed using various weightings of DLIR and ASIR-V algorithms under a low-dose DE-CTPA scanning protocol. The results demonstrated that the DLIR-H algorithm significantly outperformed DLIR-M, ASIR-V 50%, and ASIR-V 70%, highlighting its superior ability to enhance image quality under reduced radiation exposure.

Previous studies support the findings of this study. For instance, Xu et al. evaluated abdominal imaging at different keV levels (40, 50, 74, 100 keV) and concluded that 40 keV was optimal for producing high-quality abdominal aorta images using the DLIR algorithm (16). Similarly, Hosseini-Siyanaki et al. assessed VMI image quality across multiple energy levels (35–200 keV), finding that 40 keV provided the highest CT values for pulmonary arteries, although this was accompanied by increased image noise. In contrast, 76–79 keV produced the highest CNR and SNR for the pulmonary arteries (main pulmonary artery and bilateral pulmonary arteries) (17). While this study limited itself to 40 keV VMI images, it underscores the importance of determining the optimal keV level for clinical use. Therefore, future studies should compare the image quality of different VMI energy levels under low-dose DE-CTPA protocols to optimize settings for pulmonary imaging.

The mean contrast agent volume in this study was 26 mL, representing a 48% reduction compared with the conventional 50 mL dose used in standard CTPA protocols at our institution. This reduction in contrast agent volume may reduce the risk of CIN, particularly in patients with renal insufficiency or those requiring repeated imaging procedures (18). However, as compared to Zhang et al., our study saw a 50.4% increase in the effective radiation dose (19). This rise in dose can be attributed to the relatively low NI used in the imaging protocol. To mitigate this issue, future investigations should explore the use of 370 mgI/mL iodinated contrast agents and optimize scanning parameters to reduce radiation exposure without compromising image quality.

The improved image quality obtained with DLIR-H can be explained by the unique characteristics of deep learning models. Unlike traditional iterative reconstruction algorithms such as ASIR-V, which rely on noise modeling and image smoothing that may obscure fine details, DLIR utilizes CNNs. By training on large datasets of high-quality images, DLIR can effectively separate noise from anatomical details, preserving finer structures while minimizing noise. This is particularly advantageous for low-dose CT, where minimizing noise not only enhances CNR but also boosts SNR, thereby improving diagnostic accuracy and clinical decision-making (20,21).

DLIR offers three selectable strength levels (low, medium, and high), which represent different balances between noise reduction and preservation of image texture (14). DLIR-H applies stronger noise suppression, resulting in smoother images with higher CNR and SNR, whereas DLIR-M provides moderate suppression and better preserves spatial resolution and natural vascular texture. DLIR-L, while available, applies minimal noise suppression and yields image quality closer to conventional iterative reconstruction; therefore, it was not included in this study. Given the inherently high noise levels of 40 keV VMI, we focused on DLIR-M and DLIR-H as the most clinically relevant options for optimizing DE-CTPA protocols.

Deep learning-based reconstruction offers clear advantages in noise reduction and preservation of fine anatomical details compared with conventional iterative methods. However, one concern is the possibility that such models may introduce hallucinated structures or amplify artifacts due to their reliance on learned priors. In this study, no hallucinated structures or amplified artifacts were observed in DLIR reconstructions when compared with ASIR-V images. The pulmonary vascular anatomy and embolic lesion depiction remained consistent across methods, and subjective evaluations by two radiologists did not identify any DLIR-specific artifacts. Nevertheless, given the theoretical risk of hallucination inherent to deep learning models, further validation in larger multicenter studies is required.

A limitation of this study is that both ASIR-V and DLIR are vendor-specific algorithms developed by GE Healthcare. Although the general principles of iterative and deep learning reconstruction are similar across CT platforms, variations in algorithm design, training datasets, and system hardware mean that our findings may not be directly generalizable to scanners from other vendors or institutions. Validation studies across different platforms are needed to confirm the broader applicability of these results.

Additionally, the study found that 40 keV VMI images reconstructed using ASIR-V 70% provided superior image quality compared to ASIR-V 50%. This aligns with previous findings by Yan Luo et al., who demonstrated that ASIR-V 60% yielded lower SD values and higher CNR and SNR than ASIR-V 30% in CT small bowel imaging for obese patients (22). Han Zhang et al. confirmed that ASiR-V was positively correlated with CNR, while higher ASiR-V weightings (above 80%) resulted in waxy artifacts that degraded image quality (23). These findings highlight the complex relationship between ASIR-V weightings and image quality under low-dose scanning conditions. Consequently, further exploration is necessary to better understand how varying ASIR-V weightings influence VMI image quality.

A recent study on vertebral artery imaging reported that DLIR-H significantly improved the quality of 40 keV VMI under a dual-low protocol. Compared with that work, our study differs in its clinical focus on pulmonary angiography, where evaluation of pulmonary arteries is the main objective (24). Furthermore, the innovation of our protocol lies in substantial contrast dose reduction (26 vs. 34 mL in the vertebral artery study). By addressing the balance between enhanced vascular contrast and increased noise at 40 keV, our study extends the application of DLIR to pulmonary CTA and provides complementary insights to prior research on smaller-caliber vessels. This highlights the novelty of our work, demonstrating that DLIR can enable diagnostically reliable image quality in high-noise but high-contrast conditions for contrast-sparing pulmonary CTA protocols.

The findings of this study highlight the significant clinical value of the DLIR-H algorithm, especially in reducing radiation dose and contrast agent volume while maintaining high image quality. This is particularly beneficial in the diagnosis of pulmonary embolism through CTPA, where traditional methods require high radiation doses and large contrast volumes, increasing risks such as radiation-induced tissue damage and CIN. By improving image quality under low-dose protocols, DLIR-H enables safer imaging, particularly for high-risk populations such as the elderly or those with renal conditions. The algorithm’s dual capability to enhance diagnostic accuracy and minimize patient exposure to radiation establishes it as a highly promising tool in clinical practice. Furthermore, DLIR-H’s application could extend to other imaging techniques, including coronary and abdominal CT, advancing personalized medicine and improving overall patient safety.

While the study provides valuable insights, several limitations should be acknowledged. First, this study was designed to evaluate objective and subjective image quality rather than diagnostic performance. The presence of pulmonary embolism was not analyzed here, as this work did not include thrombus assessment. While improved image quality is likely to facilitate the detection of small emboli and subtle vascular abnormalities, confirmation of its clinical significance will require future studies specifically designed to assess diagnostic accuracy in pulmonary embolism. Second, the relatively small sample size limits the generalizability of these findings. Therefore, multi-center trials are needed to validate the results. Third, the influence of BMI on radiation dose and contrast agent volume was not considered. Future studies should develop individualized scanning protocols based on BMI to improve precision in radiation dose control. Additionally, the study did not account for the impact of cardiac function on CTPA image quality, which can significantly affect the results. Finally, the study focused on 40 keV VMI images, and a more comprehensive analysis comparing different keV levels would be valuable for optimizing imaging protocols for DE-CTPA.

Conclusions

In conclusion, the DLIR-H algorithm significantly enhances the image quality of 40 keV VMI images under low-dose DE-CTPA scanning protocols. It outperforms DLIR-M, ASIR-V 50%, and ASIR-V 70%, making it a promising tool for improving image quality in CTPA, particularly in clinical settings where minimizing radiation dose and contrast agent volume is essential.

Acknowledgments

None.

Footnote

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1420/dss

Funding: The study was supported by financial support from the Affiliated Hospital of Xuzhou Medical University Institutional Research Project (No. 2024ZY08) and Xuzhou Municipal Health and Health Commission Science and Technology Project (No. XWKYHT20240086).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1420/coif). A.S. is employed by GE HealthCare China. 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 Institutional Review Board of the Affiliated Hospital of Xuzhou Medical University (approval No. XYFY2023-KL316-01) and informed consent was obtained from all individual participants.

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. Kahn SR, de Wit K. Pulmonary Embolism. N Engl J Med 2022;387:45-57. [Crossref] [PubMed]
2. Meinel FG, Nance JW Jr, Schoepf UJ, Hoffmann VS, Thierfelder KM, Costello P, Goldhaber SZ, Bamberg F. Predictive Value of Computed Tomography in Acute Pulmonary Embolism: Systematic Review and Meta-analysis. Am J Med 2015;128:747-59.e2. [Crossref] [PubMed]
3. Cozzi D, Moroni C, Cavigli E, Bindi A, Caviglioli C, Nazerian P, Vanni S, Miele V, Bartolucci M. Prognostic value of CT pulmonary angiography parameters in acute pulmonary embolism. Radiol Med 2021;126:1030-6. [Crossref] [PubMed]
4. Doğan H, Kroft LJ, Huisman MV, van der Geest RJ, de Roos A. Right ventricular function in patients with acute pulmonary embolism: analysis with electrocardiography-synchronized multi-detector row CT. Radiology 2007;242:78-84. [Crossref] [PubMed]
5. Albrecht MH, Vogl TJ, Martin SS, Nance JW, Duguay TM, Wichmann JL, De Cecco CN, Varga-Szemes A, van Assen M, Tesche C, Schoepf UJ. Review of Clinical Applications for Virtual Monoenergetic Dual-Energy CT. Radiology 2019;293:260-71. [Crossref] [PubMed]
6. Wang Y, Wang Y, Hu P, Niu Z, Hu H. Comparative study of the image quality of twin beam dual energy and single energy carotid CT angiography. Eur J Radiol 2022;148:110160. [Crossref] [PubMed]
7. Greffier J, Viry A, Barbotteau Y, Frandon J, Loisy M, de Oliveira F, Beregi JP, Dabli D. Phantom task-based image quality assessment of three generations of rapid kV-switching dual-energy CT systems on virtual monoenergetic images. Med Phys 2022;49:2233-44. [Crossref] [PubMed]
8. Jiang C, Jin D, Liu Z, Zhang Y, Ni M, Yuan H. Deep learning image reconstruction algorithm for carotid dual-energy computed tomography angiography: evaluation of image quality and diagnostic performance. Insights Imaging 2022;13:182. [Crossref] [PubMed]
9. Li W, Huang W, Li P, Wen Y, Shuai T, He Y, You Y, Yu J, Diao K, Song B. Application of deep learning image reconstruction-high algorithm in one-stop coronary and carotid-cerebrovascular CT angiography with low radiation and contrast doses. Quant Imaging Med Surg 2024;14:1860-72. [Crossref] [PubMed]
10. Ichikawa Y, Kanii Y, Yamazaki A, Nagasawa N, Nagata M, Ishida M, Kitagawa K, Sakuma H. Deep learning image reconstruction for improvement of image quality of abdominal computed tomography: comparison with hybrid iterative reconstruction. Jpn J Radiol 2021;39:598-604. [Crossref] [PubMed]
11. Park C, Choo KS, Jung Y, Jeong HS, Hwang JY, Yun MS. CT iterative vs deep learning reconstruction: comparison of noise and sharpness. Eur Radiol 2021;31:3156-64. [Crossref] [PubMed]
12. Abdellatif W, Esslinger E, Kobes K, Wong A, Powell J, Ali IT, Andrews G, Nicolaou S. Acquisition time, radiation dose, subjective and objective image quality of dual-source CT scanners in acute pulmonary embolism: a comparative study. Eur Radiol 2020;30:2712-21. [Crossref] [PubMed]
13. Pouget E, Dedieu V, Magnin ML, Biard M, Lienemann G, Garcier JM, Magnin B. Response surface methodology for predicting optimal conditions in very low-dose chest CT imaging. Phys Med 2025;131:104916. [Crossref] [PubMed]
14. Thibault JB, Nett B, Tang J, Liu E. TrueFidelity™ for Gemstone™ Spectral Imaging: A new generation of spectral imaging powered by deep learning. Available online: https://www.gehealthcare.com/-/jssmedia/gehc/us/images/products/computed-tomography/truefidelity/related-content/new/truefidelity-for-gsi-whitepaper_digital_jb19879xx_final.pdf?rev=-1. Accessed May 2022.
15. Geyer LL, Schoepf UJ, Meinel FG, Nance JW Jr, Bastarrika G, Leipsic JA, Paul NS, Rengo M, Laghi A, De Cecco CN. State of the Art: Iterative CT Reconstruction Techniques. Radiology 2015;276:339-57. [Crossref] [PubMed]
16. Xu JJ, Lönn L, Budtz-Jørgensen E, Hansen KL, Ulriksen PS. Quantitative and qualitative assessments of deep learning image reconstruction in low-keV virtual monoenergetic dual-energy CT. Eur Radiol 2022;32:7098-107. [Crossref] [PubMed]
17. Hosseini-Siyanaki M, Sagdic HS, Raviprasad AG, Munjerin SE, Prodigios JC, Anthony EY, Hochhegger B, Forghani R. Multi-Energy Evaluation of Image Quality in Spectral CT Pulmonary Angiography Using Different Strength Deep Learning Spectral Reconstructions. Acad Radiol 2025;32:2953-65. [Crossref] [PubMed]
18. Hennessey B, Danenberg H, De Vroey F, Kirtane AJ, Parikh M, Karmpaliotis D, Messenger JC, Strobel A, Curcio A, van Mourik MS, Eshuis P, Escaned J. Dynamic Coronary Roadmap versus standard angiography for percutaneous coronary intervention: the randomised, multicentre DCR4Contrast trial. EuroIntervention 2024;20:e198-206. [Crossref] [PubMed]
19. Zhang X, Huang H, Huang Y, Sun Y, Chen Z, Chen M, Xu J. Reduction of radiation dose and contrast medium volume in computed tomography pulmonary angiography: adaptation of dual-energy computed tomography (CT) protocols to the body mass index. Clin Radiol 2025;86:106944. [Crossref] [PubMed]
20. Li H, Zhang Y, Hua S, Sun R, Zhang Y, Yang Z, Peng Y, Sun J. Deep Learning Image Reconstruction (DLIR) Algorithm to Maintain High Image Quality and Diagnostic Accuracy in Quadruple-low CT Angiography of Children with Pulmonary Sequestration: A Case Control Study. Acad Radiol 2025;32:4197-205. [Crossref] [PubMed]
21. Zhu H, Huang Z, Chen Q, Ma W, Yu J, Wang S, Tao G, Xing J, Jiang H, Sun X, Liu J, Yu H, Zhu L. Feasibility of Sub-milliSievert Low-dose Computed Tomography with Deep Learning Image Reconstruction in Evaluating Pulmonary Subsolid Nodules: A Prospective Intra-individual Comparison Study. Acad Radiol 2025;32:2309-19. [Crossref] [PubMed]
22. Luo Y, Tang Y, Chen S, Tang H, Cheng Y, Zhang F, Shen Y, Wang Q. Improving image quality and diagnostic performance using deep learning image reconstruction in 100-kVp CT enterography for patients with wide-range body mass index. Eur J Radiol 2025;189:112167. [Crossref] [PubMed]
23. Zhang H, Han H, Li Y, Yang Q, Yin T, Yu W. Effects of different combinations of noise index and preset adaptive statistical iterative reconstruction Veo on the accuracy and image quality of bone mineral density measurements using fast kilovolt-switching dual-energy computed tomography: a phantom study. Quant Imaging Med Surg 2025;15:164-76. [Crossref] [PubMed]
24. Tang Y, Zhang H, Chen L, Yu M, Zhang H, Zhang D, Wu Y, Liu Z, Sun A, Meng Y, Xu K. Comparison of image quality of 40 keV virtual monoenergetic images of vertebral arteries using DLIR and ASIR-V algorithms under a dual-low scanning protocol. Eur J Radiol 2025;191:112276. [Crossref] [PubMed]