Deep learning precise image reconstruction algorithm for abdominal CT: impact on image quality and radiation dose reduction—a model validation study

Introduction

Computed tomography (CT) is widely used for the screening and evaluation of various diseases across different anatomical regions due to its rapid and non-invasive nature (1,2). However, in abdominal CT imaging, minimizing the radiation dose while adhering to the as low as reasonably achievable (ALARA) principle remains a persistent challenge. Previous studies have emphasized the importance of appropriate indication selection to reduce the overall number of abdominal CT examinations and recommended optimizing scanning protocols to mitigate radiation-induced harm to patients (3,4). One of the proven and effective strategies to reduce radiation dose in CT imaging is the use of iterative reconstruction (IR) algorithms (5,6). These algorithms reduce image noise in low-dose CT scans, helping to maintain acceptable image quality. In clinical practice, the noise-reduction capability of these algorithms is often translated into dose reductions to preserve comparable noise levels. However, the drawbacks of IR images, such as excessive smoothing, wax-like appearance (7), and long reconstruction time (8), remain issues that the industry is striving to overcome. In recent years, artificial intelligence reconstruction algorithms based on deep neural networks (DNNs) have been proposed and begun to be applied in CT image reconstruction to address the challenges faced by IR algorithms (9). Currently, there are two main types of deep learning-based artificial intelligence algorithms widely used in CT: TrueFidelity (10) developed by GE Healthcare Systems (Chicago, IL, USA) and AiCE (11) developed by Canon Medical Systems Corporation (Otawara, Japan). The former utilizes filtered back projection (FBP) datasets, while the latter relies on model-based IR datasets. The integration of these reconstruction algorithms holds the potential to effectively distinguish between signals and noise, thereby reducing noise while preserving image texture as much as possible. Numerous published studies have demonstrated that these deep learning reconstruction (DLR) algorithms represent a groundbreaking advancement compared to the IR algorithms used so far in terms of dose optimization and image quality improvement (12-16). Therefore, exploring the potential of DLR algorithms to reduce CT scan dose while lowering image noise levels and enhancing image quality holds new guiding significance for the formulation of low-dose abdominal CT scanning protocols.

The Philips deep learning precise image (DLPI) algorithm has recently been introduced as a DNN-based reconstruction engine, DLPI employs convolutional neural networks (CNNs) trained on high-quality FBP datasets to distinguish signal from noise and intelligently suppress noise without compromising anatomical or pathological detail (17). Therefore, this study aims to retrospectively validate a series of abdominal CT images reconstructed using the newly developed DLPI algorithm, and to evaluate and verify the application of this DLPI algorithm in improving abdominal image quality and detail visualization capability. Additionally, by altering slice thickness to simulate changes in X-ray flux, this study intends to predict the DLPI algorithm’s dose reduction potential under prospective low-dose scanning conditions.

Methods

Patient information

This retrospective study initially enrolled 107 patients who were admitted to Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, from October 8, 2024 to January 31, 2025, but only 60 patients were ultimately included (the study flowchart is shown in Figure 1). The inclusion criteria were as follows: (I) clinical diagnosis of abdominal trauma or other abdominal diseases; (II) patients with good cooperation in breath-holding and no motion artifacts in their CT images; (III) patients with complete clinical data and follow-up records; and (IV) age >18 years. The exclusion criteria were as follows: (I) patients whose raw data had been overwritten and were unavailable; (II) patients with poor cooperation in breath-holding, resulting in motion artifacts in their CT images. This study was not registered on any platform due to it being a small-sample exploratory study of imaging technology. All abdominal CT scans were performed in accordance with hospital-approved guidelines and regulations. and included patients who presented with abdominal trauma, acute abdominal pain, or other symptoms, or those referred by clinicians for the exclusion of abdominal abnormalities.

Figure 1 Flowchart of the inclusion and exclusion of the study population.

Scanning parameters

CT scans were performed using a 64-slice multidetector CT scanner (Incisive CT, Philips Healthcare, Amsterdam, the Netherlands). All patients underwent helical scanning with breath-hold after inspiration, following respiratory instructions. The scanning protocol employed a tube voltage of 120 kVp and a gantry rotation time of 0.8 s. Automatic tube current modulation was applied with the Dose Right Index (DRI) set to 18, and the tube current range limited to 80–200 mA. The raw data were reconstructed into different slice thicknesses (5 and 1 mm) using the FBP, iDose4, and DLPI algorithms, respectively. iDose4 images were generated using the standard level 4 weighting of the iDose IR algorithm and served as the clinical reference standard. DLPI images were reconstructed using the high-weight mode (PI-smooth) to further reduce image noise (18). The volume CT dose index (CTDIvol) and dose-length product (DLP) were recorded for each patient.

Image quality evaluation

Subjective evaluation

All images were transferred to the post-processing workstation (Philips IntelliSpace Portal 10.0, Philips Healthcare). During the evaluation, all patient-identifying information and reconstruction parameters were anonymized. Observers were allowed to adjust the window width and window level as appropriate. Two abdominal radiologists, with 15 and 10 years of diagnostic experience respectively, independently assessed the image quality based on a predefined scoring system. In cases of disagreement, consensus was reached through discussion; if no agreement could be made, a senior abdominal radiologist with 25 years of diagnostic experience made the final decision. Subjective image quality was scored using an improved 5-point scale (19), where a score of 3 represented diagnostically acceptable image quality. The evaluation focused on the clarity of intestinal loop, the delineation of boundaries between abdominal organs, and the overall image quality. A score of 1 indicated non-diagnostic images with severe noise, non-visualization of intestinal loop, and indistinguishable organ boundaries. A score of 2 reflected barely detectable but still sub-diagnostic images with significant noise, poorly visualized intestinal loop, and unclear organ margins. A score of 3 corresponded to diagnostic images with moderate noise and slightly blurred but measurable boundaries. A score of 4 denoted good diagnostic quality with low noise, clearly identifiable intestinal loop, and well-defined organ boundaries. A score of 5 represented excellent image quality with virtually no noise, sharply defined intestinal loop, and clear delineation of abdominal structures. Overall image quality was primarily rated according to the radiologist’s confidence in making a diagnosis (the specific scoring criteria are detailed in Table 1). Additionally, the number of lesions detected in each reconstruction group was recorded, including hepatic cysts, splenic hematomas, bowel obstructions, renal calculi, peritonitis, and abdominal masses. The number of lesions was recorded for each patient based on clear lesion margins, asymmetry of abdominal organs, and abnormal tissue densities.

Objective evaluation

Two radiologists jointly performed the objective quantitative measurements on the workstation. For each patient, regions of interest (ROIs) were placed at the level of the liver, spleen, abdominal aorta, and pancreas to measure the CT values and standard deviations (SDs) of these organs. Specifically, using the copy-and-paste function, ROIs were placed in homogeneous areas at the center of the liver, as well as on the spleen, pancreas, abdominal aorta, and erector spinae muscle at the same axial level. The mean CT values and corresponding SDs for the liver, spleen, pancreas, abdominal aorta, and psoas muscle were recorded. The signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were calculated using the following formulas:

SNR(tissue)=CTvalue(tissue)/SDvalue(erectorspinaemuscle)[1]

CNR(tissue)=(CTvalue(tissue)−CTvalue(erectorspinaemuscle))/SDvalue(erectorspinaemuscle)[2]

Statistical analysis

Statistical analysis was performed using SPSS version 26.0. Continuous variables were expressed as mean ± SD. For the objective evaluation, normality of the data (including CT values, SD, SNR, and CNR) across different reconstruction algorithms at the same slice thickness was tested using the normal distribution test. For variables following a normal distribution, repeated measures analysis of variance (ANOVA) was used, with Tukey’s honestly significant difference (HSD) post hoc tests applied for pairwise comparisons to identify specific group differences. For ordinal or non-normally distributed variables, the Friedman test was applied, followed by Dunn’s post hoc tests with Bonferroni correction to further locate significant differences between groups. In this study, we first analyzed the differences in each region of interest among different reconstruction algorithms at the same slice thickness, Subsequently, we compared 5 mm iDose4 images with 1 mm PI-smooth images to assess the potential of the DLPI algorithm to reduce radiation dose while maintaining image noise levels and diagnostic acceptability equivalent to 5 mm reconstructions. A P value <0.05 was considered statistically significant. For subjective evaluation, the median score was used to assess image quality across reconstruction groups. The Friedman test was used for ordinal data analysis, and a P value <0.01 was considered statistically significant. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was reviewed and approved by the Institutional Review Board of Union Hospital, Tongji Medical College, Huazhong University of Science and Technology (grant No. UHCT-IEC-SOP-016-03-01), which granted a waiver of informed consent due to the retrospective and anonymized nature of the data.

Results

General information statistics

A total of 60 patients were enrolled in this study, including 35 males and 25 females. The mean age of the study cohort was 53.58±13.24 years, and the mean body mass index (BMI) was 25.21±2.71 kg/m². Regarding radiation dose parameters, the mean tube current, CTDIvol, and DLP were 87.13±21.63 mAs, 7.10±1.71 mGy, and 308.18±128.46 mGy·cm, respectively. Regarding clinical indications, 32 patients (53.3%) presented with acute abdominal pain, 14 patients (23.3%) with abdominal trauma, and another 14 patients (23.3%) with other symptoms. For lesion characteristics, 48 patients (80.0%) had abnormal findings, while 12 patients (20.0%) had no abnormal findings; the total lesion count per patient had a mean of 1.8±0.9 and a median of 2 (range, 0–4) (see Table 2 for details).

Subjective evaluation results

Two radiologists conducted a blinded evaluation of the image quality of six image sets with different algorithms and slice thicknesses. They were unaware of the reconstruction parameters of the six image sets. During the evaluation, there were discrepancies in the scores of four cases, which were ultimately resolved through discussion after review by a radiologist with 25 years of experience in abdominal diagnosis. The subjective scoring results are summarized in Table 3. As expected, under low-dose scanning conditions, FBP reconstruction at the standard 5 mm slice thickness failed to fully meet diagnostic requirements due to its relatively high image noise, receiving a subjective score of 2. In contrast, both iDose4 and PI-smooth reconstructions at 5 mm slice thickness achieved acceptable noise levels and met diagnostic standards, each receiving a score of 4. The PI-smooth algorithm consistently received the highest subjective ratings in both the 5 and 1 mm reconstructions, with a score of 4 in both cases.

Significant differences were observed between iDose4 and PI-smooth images in terms of intestinal loop delineation, inter-organ boundary clarity, and overall image quality (P<0.001). Among the 1 mm slice thickness reconstructions, only the PI-smooth images met diagnostic criteria, scoring 4 points. Additionally, there was no statistically significant difference in subjective image quality scores between the 1 mm PI-smooth and the 5 mm iDose4 images (P=0.39). Representative images for each algorithm and slice thickness are illustrated in Figure 2.

Figure 2 Comparison of image quality between different reconstruction algorithms at 5 and 1 mm slice thicknesses. The illustrated patient is a 58-year-old male with a BMI of 26.4 kg/m² who presented with several days of upper abdominal discomfort and underwent a low-dose abdominal CT scan as requested by the clinician. (A-F) Represents the 5 mm slice thickness group: (A,D) were reconstructed using FBP, showing a round hypodense lesion in the left hepatic lobe; however, the overall image noise was relatively high, and although the details of abdominal organs and intestinal loop was moderately visible, the subjective image quality score was 3. (B,E) and (C,F) were reconstructed using iDose4 and PI-smooth, respectively; both exhibited lower overall noise, with better visualization of abdominal structures, and the hypodense hepatic lesion appeared clearly with sharp margins, resulting in a subjective score of 4. (G-L) Represents the 1 mm slice thickness group: (G,J) were reconstructed with FBP, while (H,K) and (I,L) were reconstructed with iDose4 and PI-smooth, respectively. Compared with the 5 mm images, the reduced slice thickness in (G,J) and (H,K) resulted in increased image noise, which impaired the visualization of abdominal organ morphology and boundaries, and the lesion contour in the liver was poorly defined, making the images non-diagnostic. In contrast, (I,L) showed substantially reduced noise, with better depiction of intra-abdominal structural details; the round hypodense lesion in the left hepatic lobe was clearly visible with sharp margins, and the overall image quality was comparable to that of the 5 mm iDose4 reconstruction, also receiving a subjective score of 4. BMI, body mass index; CT, computed tomography; FBP, filtered back projection.

Two radiologists identified a total of 65 lesions in 60 patients using 5-mm slice thickness images reconstructed with three algorithms. These lesions included 15 hepatic cysts, 10 splenic hematomas, 5 cases of intestinal obstruction, 10 renal calculi, 10 cases of peritonitis, and 15 abdominal masses. Compared with 5 mm slice thickness images, 1 mm slice thickness images observed an additional 3—7 small hepatic cysts and splenic hematomas, with this effect being most significant in 1-mm PI-smooth images (with an increase of 7 cases) (see Table 4 for details). Although the lesion counts represent preliminary observations by the two radiologists, the numerical data still reflect that 1-mm slice thickness images from the PI-smooth algorithm demonstrated greater advantages in both overall image quality scores and detail visualization (Tables 3,4).

Objective evaluation results

Two experienced diagnostic radiologists measured, analyzed, and calculated the CT values, SD, SNR, and CNR for six groups of images. Statistical analysis confirmed that all data followed a normal distribution. Repeated measures ANOVA results showed no significant differences in CT values across all ROIs among the groups (all P>0.05), while differences in SD, SNR, and CNR were statistically significant (all P<0.001), as detailed in Table 5.

Post-hoc pairwise comparisons showed that, at both 5 and 1 mm slice thicknesses, PI-smooth consistently achieved the lowest SD and the highest SNR and CNR across all organs. Taking the liver as an example: at 5 mm, PI-smooth SD was significantly lower than that of FBP and iDose4 (both P<0.001, Cohen’s d=3.89 and 3.84, respectively); the same pattern was observed at 1 mm (vs. FBP: P<0.001, d=5.96; vs. iDose4: P<0.001, d=3.81). Overall, values obtained with 5 mm thickness were higher than those with 1 mm; for instance, liver SNR in the 5 mm PI-smooth group exceeded that in the 1 mm PI-smooth group (9.04±2.58 vs. 5.93±1.56, P<0.001, d=1.94), as did liver CNR (P<0.001, d=0.61).

Notably, SNR of the 1 mm PI-smooth images was significantly higher than that of the 5 mm iDose4 images in every organ evaluated (liver: P<0.001, d=1.83; spleen: P<0.001, d=1.94; abdominal aorta: P<0.001, d=1.85; pancreas: P<0.001, d=1.86; erector spinae muscle: P<0.001, d=1.91).

Discussion

In this study, we evaluated the performance of the DLPI algorithm in enhancing image quality, particularly demonstrating significant improvements compared to the conventional hybrid IR iDose4 algorithm at the same slice thickness. By reducing the slice thickness to simulate a decrease in X-ray flux, DLPI was able to maintain image noise levels comparable to those of conventional thicker slices (5 mm) even at thinner slices (1 mm). These findings suggest that, while maintaining the current diagnostic standard slice thickness of 5 mm, the use of DLPI enables potential radiation dose reduction and improves lesion detection capability through thinner slice imaging.

Abdominal CT is widely used for the diagnosis of abdominal trauma, acute abdominal pain, and space-occupying lesions due to its efficiency (20), However, with its increasing utilization, concerns about radiation exposure have intensified, leading to a growing demand for dose reduction strategies (21,22). One effective approach to reducing CT radiation dose is the application of IR algorithms (23,24). IR algorithms can enhance image quality under low signal conditions, thereby enabling dose reduction. Several validated IR techniques include Adaptive Statistical Iterative Reconstruction (ASIR-V) by GE Healthcare, Advanced Modeled Iterative Reconstruction (ADMIRE) by Siemens (Erlangen,Germany), and the Iterative Model Reconstruction (IMR), all of which have been reported to improve image quality in abdominal CT (25-27). In the present study, we reconstructed both thin- and thick-slice images using FBP, iDose4, and high-strength DLPI (PI-smooth) algorithms to evaluate the differences among reconstruction techniques. As demonstrated in previous studies, our findings confirm that the iDose4 algorithm can improves liver SNR by 24.3% in 5 mm slices and by 43.4% in 1 mm slices. Moreover, our results indicate that PI-smooth can further enhance image quality by substantially reducing noise in both thin- and thick-slice images. Compared to conventional FBP reconstruction, PI-smooth increased liver SNR by 156.0% at 5 mm slice thickness and by 223.0% at 1 mm slice thickness (the specific data are shown in Table 5).

We further compared the performance of PI-smooth with that of iDose4. Although both algorithms achieved a subjective score of 4 points in the 5 and 1 mm slice thickness groups, PI-smooth in 1-mm images had a significantly higher SNR than iDose4 in 5-mm images (P<0.001). Additionally, the 1-mm PI-smooth images provided richer details, which was more conducive to radiologists’ detection of small lesions (see Table 4). For iDose4 reconstruction, only level 4 was selected, as higher levels tend to produce a blotchy appearance and blurred margins (28), which we believe is related to the intrinsic design of current IR algorithms. Although the underlying principles of IR algorithms differ across manufacturers, plastic-like artifacts remain an issue, and these artifacts become more pronounced as the IR weight increases, negatively impacting subjective image assessment. In a recent phantom study, Greffier et al. (29) found that the iDose4 algorithm reduces the peak of the noise power spectrum (NPS) during image reconstruction, particularly at higher weight levels. However, this also suppresses high-frequency signals, leading to a reduction in the average spatial frequency of the NPS at higher IR weights. In contrast, the new DLPI algorithm (17) represents a novel reconstruction technique that differs from traditional methods, consisting of three key components: data domain filtering, hybrid IR, and a DNN-based restoration module. The training process involves DNN-based inference on the dataset, which is then matched with an ideal FBP image. During reconstruction, this algorithm reduces noise levels and suppresses artifact generation while preserving the noise texture of FBP images (18,30). This approach does not compromise the original noise texture of the CT images or the radiologist’s ability to assess anatomical or pathological structures, thereby offering significant clinical advantages (31), as confirmed by our study.

Due to the limitation of scanning each patient only once to minimize radiation exposure, it is typically challenging to directly correlate the noise-reduction capability of the DLPI algorithm with its dose-reduction potential. In this study, we simulated high- and low-dose imaging by comparing images with different slice thicknesses. Since the X-ray flux used for image reconstruction is approximately proportional to the slice thickness, 1 mm images utilize only about one-fifth of the signal compared to 5 mm images, representing an estimated 80% dose reduction. However, our results demonstrated that even at one-fifth the signal strength, the SNR of 1 mm PI-smooth images remained higher than that of 5 mm iDose4 images (see Table 5), and the subjective image quality scores of both image types were comparable (both scored 4 points, see Table 3). In the present study, the image noise level and quality of 5 mm iDose4 images and 1 mm PI-smooth images were both deemed clinically acceptable for routine diagnostic use. These findings indicate that using the PI-smooth algorithm with a 5 mm slice thickness holds promise for prospectively reducing the scan dose by 80%, while still meeting diagnostic requirements and maintaining or even exceeding the SNR of the current 5 mm protocol. Naturally, if slice thickness is reduced further below 1 mm, the achievable dose reduction potential will decrease accordingly. In clinical practice, thinner slices are often preferred, as they improve spatial resolution and provide more detailed anatomical information, thereby enhancing lesion detectability. Thin-slice images can more clearly depict subtle anatomical structures such as small hepatic vessels and lesions, the layered structure of the intestinal wall, and the margins and internal architecture of soft-tissue masses. This enhanced visualization facilitates early detection and diagnosis, offering more accurate support for clinical decision-making and treatment planning.

Our study can be improved in several aspects. First, due to the limitation that DLPI reconstruction could only be performed using raw data stored on the Incisive CT, the final sample size was relatively small (n=60). Second, this was a retrospective study in which the same raw datasets were reconstructed into images with different slice thicknesses (5 and 1 mm) using different algorithms, and the signal reduction was simulated based on the known relationship between slice thickness and X-ray flux. Although we demonstrated that DLPI (PI-smooth) could achieve comparable image quality scores and higher SNR at one-fifth the slice thickness compared to current iDose4 reconstructions, we did not prospectively reduce the radiation dose during patient scanning. Further prospective studies are needed to evaluate image quality under actual dose-reduction protocols. Third, our study solely discusses the impact of Philips’ DLPI algorithm on the image quality of reconstructed images with different thicknesses, thereby inferring the potential of DLPI for dose reduction. However, there are differences between algorithms from different manufacturers, and this study does not involve a horizontal comparison between DLPI algorithms of devices from other manufacturers. In future studies on DLPI, we plan to incorporate algorithms from other manufacturers for further comparison to improve the accuracy of conclusions. Finally, since the CT is installed in a physical examination center, all included cases were non-contrast scans, and contrast-enhanced studies were not assessed. Therefore, the performance of DLPI in contrast-enhanced imaging and angiographic applications, particularly in terms of image quality enhancement and dose reduction potential, requires further investigation.

Conclusions

Compared with the iDose4 algorithm, the deep learning-based artificial intelligence DLPI algorithm can improve the image quality of thin-slice abdominal CT scans. Specifically, 1-mm slice thickness PI-smooth images maintain the same noise levels and subjective scores as 5-mm slice thickness iDose4 images, while simultaneously enhancing the ability to display image details. Furthermore, the X-ray flux required for 1-mm slice thickness is only one-fifth of that for 5-mm slice thickness. This indicates that when maintaining the current diagnostic image quality and slice thickness standard—with 5-mm slice thickness iDose4 images as the quality benchmark—the application of the PI-smooth algorithm is expected to achieve a nearly 80% reduction in radiation dose in future prospective scanning protocols. Thus, it is worthy of clinical application and further exploration.

Acknowledgments

We appreciate the contributions of all authors for participating in data analysis, drafting, and critical revisions.

Footnote

Data Sharing Statement: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-173/dss

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-173/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-173/coif). The 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 protocol was reviewed and approved by the Institutional Review Board of Union Hospital, Tongji Medical College, Huazhong University of Science and Technology (grant No. UHCT-IEC-SOP-016-03-01), which granted a waiver of informed consent due to the retrospective and anonymized nature of the data.

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. Smith-Bindman R, Lipson J, Marcus R, et al. Radiation dose associated with common computed tomography examinations and the associated lifetime attributable risk of cancer. Arch Intern Med 2009;169:2078-86. [Crossref] [PubMed]
2. Pandharipande PV, Reisner AT, Binder WD, et al. CT in the Emergency Department: A Real-Time Study of Changes in Physician Decision Making. Radiology 2016;278:812-21. [Crossref] [PubMed]
3. Ferorelli D, Donno F, De Giorgio G, et al. Head CT scan in emergency room: Is it still abused? Quantification and causes analysis of overprescription in an Italian Emergency Department. Radiol Med 2020;125:595-9. [Crossref] [PubMed]
4. Strait L, Sussman R, Ata A, et al. Utilization of CT imaging in minor pediatric head, thoracic, and abdominal trauma in the United States. J Pediatr Surg 2020;55:1766-72. [Crossref] [PubMed]
5. Nagayama Y, Nakaura T, Tsuji A, et al. Cerebral bone subtraction CT angiography using 80 kVp and sinogram-affirmed iterative reconstruction: contrast medium and radiation dose reduction with improvement of image quality. Neuroradiology 2017;59:127-34. [Crossref] [PubMed]
6. Nagayama Y, Nakaura T, Tsuji A, et al. Radiation dose reduction using 100-kVp and a sinogram-affirmed iterative reconstruction algorithm in adolescent head CT: Impact on grey-white matter contrast and image noise. Eur Radiol 2017;27:2717-25. [Crossref] [PubMed]
7. Mileto A, Zamora DA, Alessio AM, et al. CT Detectability of Small Low-Contrast Hypoattenuating Focal Lesions: Iterative Reconstructions versus Filtered Back Projection. Radiology 2018;289:443-54. [Crossref] [PubMed]
8. Euler A, Solomon J, Marin D, et al. A Third-Generation Adaptive Statistical Iterative Reconstruction Technique: Phantom Study of Image Noise, Spatial Resolution, Lesion Detectability, and Dose Reduction Potential. AJR Am J Roentgenol 2018;210:1301-8. [Crossref] [PubMed]
9. Willemink MJ, Noël PB. The evolution of image reconstruction for CT-from filtered back projection to artificial intelligence. Eur Radiol 2019;29:2185-95. [Crossref] [PubMed]
10. Hsieh J, Liu E, Nett B, et al. A new era of image reconstruction: TrueFidelityTM Technical white paper on deep learning image reconstruction. GE Healthcare; 2019. Available online: https://www.gehealthcare.com/-/jssmedia/040dd213fa89463287155151fdb01922.pdf
11. Boedeker K. AiCE Deep Learning Reconstruction: Bringing the Power of Ultra-High Resolution CT to Routine Imaging. Technical white paper. Canon Medical Systems; 2019. Available online: https://global.medical.canon/publication/ct/2019WP_AiCE_Deep_Learning.html
12. Wang H, Yue S, Liu N, et al. Deep learning reconstruction vs standard reconstruction for abdominal CT: the influence of BMI. Eur Radiol 2024;34:1614-23. [Crossref] [PubMed]
13. Nakamoto A, Onishi H, Ota T, et al. Contrast-enhanced thin-slice abdominal CT with super-resolution deep learning reconstruction technique: evaluation of image quality and visibility of anatomical structures. Jpn J Radiol 2025;43:445-54. [Crossref] [PubMed]
14. Kazimierczak W, Kazimierczak N, Wilamowska J, et al. Enhanced visualization in endoleak detection through iterative and AI-noise optimized spectral reconstructions. Sci Rep 2024;14:3845. [Crossref] [PubMed]
15. Akagi M, Nakamura Y, Higaki T, et al. Deep learning reconstruction of equilibrium phase CT images in obese patients. Eur J Radiol 2020;133:109349. [Crossref] [PubMed]
16. Franck C, Zhang G, Deak P, et al. Preserving image texture while reducing radiation dose with a deep learning image reconstruction algorithm in chest CT: A phantom study. Phys Med 2021;81:86-93. [Crossref] [PubMed]
17. Philips Computed Tomography. White Paper: AI for Significantly Lower Dose and Improved Image Quality – Precise Image. Philips Healthcare; 2021. Available online: https://www.philips.com/c-dam/b2bhc/master/resourcecatalog/landing/precisesuite/incisive_precise_image.pdf
18. Greffier J, Si-Mohamed S, Frandon J, et al. Impact of an artificial intelligence deep-learning reconstruction algorithm for CT on image quality and potential dose reduction: A phantom study. Med Phys 2022;49:5052-63. [Crossref] [PubMed]
19. Nagata M, Ichikawa Y, Domae K, et al. Application of Deep Learning-Based Denoising Technique for Radiation Dose Reduction in Dynamic Abdominal CT: Comparison with Standard-Dose CT Using Hybrid Iterative Reconstruction Method. J Digit Imaging 2023;36:1578-87. [Crossref] [PubMed]
20. Alshamari M, Norrman E, Geijer M, et al. Diagnostic accuracy of low-dose CT compared with abdominal radiography in non-traumatic acute abdominal pain: prospective study and systematic review. Eur Radiol 2016;26:1766-74. [Crossref] [PubMed]
21. Miller DL, Vano E, Rehani MM. Reducing radiation, revising reference levels. J Am Coll Radiol 2015;12:214-6. [Crossref] [PubMed]
22. Héliou R, Normandeau L, Beaudoin G. Towards dose reduction in CT: patient radiation dose assessment for CT examinations at university health center in Canada and comparison with national diagnostic reference levels. Radiat Prot Dosimetry 2012;148:202-10. [Crossref] [PubMed]
23. Fletcher JG, DeLone DR, Kotsenas AL, et al. Evaluation of Lower-Dose Spiral Head CT for Detection of Intracranial Findings Causing Neurologic Deficits. AJNR Am J Neuroradiol 2019;40:1855-63. [Crossref] [PubMed]
24. Higaki T, Nakamura Y, Fukumoto W, et al. Clinical application of radiation dose reduction at abdominal CT. Eur J Radiol 2019;111:68-75. [Crossref] [PubMed]
25. Mørup SD, Stowe J, Kusk MW, et al. The impact of ASiR-V on abdominal CT radiation dose and image quality - A phantom study. J Med Imaging Radiat Sci 2022;53:453-9. [Crossref] [PubMed]
26. Kataria B, Nilsson Althén J, et al. Image quality and potential dose reduction using advanced modeled iterative reconstruction (ADMIRE) in abdominal CT - a review. Radiat Prot Dosimetry 2021;195:177-87. [Crossref] [PubMed]
27. Hettinger E, Aurumskjöld ML, Sartor H, et al. Evaluation of model-based iterative reconstruction in abdominal computed tomography imaging at two different dose levels. Radiat Prot Dosimetry 2021;195:205-11. [Crossref] [PubMed]
28. Sun J, Li H, Wang B, et al. Application of a deep learning image reconstruction (DLIR) algorithm in head CT imaging for children to improve image quality and lesion detection. BMC Med Imaging 2021;21:108. [Crossref] [PubMed]
29. Greffier J, Hamard A, Pereira F, et al. Image quality and dose reduction opportunity of deep learning image reconstruction algorithm for CT: a phantom study. Eur Radiol 2020;30:3951-9. [Crossref] [PubMed]
30. Si-Mohamed S, Boccalini S, Rodesch PA, et al. Feasibility of lung imaging with a large field-of-view spectral photon-counting CT system. Diagn Interv Imaging 2021;102:305-12. [Crossref] [PubMed]
31. Barca P, Domenichelli S, Golfieri R, et al. Image quality evaluation of the Precise image CT deep learning reconstruction algorithm compared to Filtered Back-projection and iDose(4): a phantom study at different dose levels Phys Med 2023;106:102517. [Crossref] [PubMed]