The application of deep learning-based artificial intelligence algorithms combined with low-dose scanning protocols in chest CT

Introduction

In recent years, an increasing awareness of the need to reduce the radiation dose in computed tomography (CT) examinations has spurred the development of dose-reduction methods (1,2). A combination of techniques (e.g., reducing the tube voltage and current, and incorporating iterative reconstruction algorithms) is commonly used to achieve radiation dose reduction (3). For instance, GE’s Adaptive Statistical Iterative Reconstruction technology not only significantly reduces the CT radiation dose but also has notable advantages over traditional filtered back projection (FBP) techniques (4). Similarly, Philips Healthcare’s iterative reconstruction technologies, such as iDose4 and iterative model reconstruction, have demonstrated excellent clinical performance, maintaining high image quality while substantially lowering the radiation dose (5). However, challenges such as excessive smoothing and a wax-like appearance in iterative reconstruction images (6), as well as prolonged reconstruction times (7), remain issues that need to be addressed.

With the widespread application of artificial intelligence (AI), new reconstruction algorithms based on deep learning reconstruction (DLR) have broken through the limitations of traditional iterative algorithms in terms of image quality (8). These algorithms improve image quality while reducing image noise and computational load, and have been widely applied in clinical practice (9,10). Over the past few years, two relatively mature DLR algorithms have been used in clinical practice: TrueFidelity, developed by GE Healthcare (Chicago, IL, USA) (11), and the Advanced Intelligent Clear-IQ Engine (AiCE), developed by Canon Medical Systems (Otawara, Japan) (12). Both algorithms use deep neural networks trained on high-quality datasets; TrueFidelity is trained using FBP datasets, while the AiCE is trained using model-based iterative reconstruction datasets (11,12). These algorithms enable the differentiation between signal and noise, effectively reducing noise while preserving image texture as much as possible. Numerous published studies have shown the effectiveness of these DLR algorithms in dose optimization and image quality improvement in the abdomen and chest (13-18). Recent research has also explored the application of DLR algorithms in other anatomical regions, such as the head and lumbar spine, showing promising results in reducing the radiation dose without compromising image quality (19,20). These studies highlight the broader potential of DLR algorithms across different imaging systems.

Recently, Philips Healthcare has developed a new deep learning-based AI reconstruction algorithm known as precise imaging (PI). Unlike other methods, this algorithm uses low-dose CT data as input, and matches that data with routine and even high-dose CT data as reference information. This approach covers imaging indicators under various scanning conditions, effectively reducing the radiation dose and artifact generation. Currently, there are limited reports on the application of this algorithm in chest imaging. Therefore, this study integrated PI with a low-dose chest CT scanning protocol for health screening to evaluate its effect on image quality. We present this article in accordance with the STROBE reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-685/rc).

Methods

Patient data

A retrospective analysis was conducted of 100 patients (66 male and 34 female) who underwent chest CT scans at the Health Examination Center of Wuhan Union Hospital between October and December 2024. The patients were divided into two groups: the experimental group (n=50), which underwent low-dose scanning, with images reconstructed using both the PI reconstruction algorithm and the iDose iterative reconstruction algorithm; and the control group (n=50), which underwent conventional-dose scanning, with images reconstructed using the iDose iterative reconstruction technique. Patients were included in the study if they met the following inclusion criteria: were aged 18 years or older, and were able to cooperate during the examination. Patients were excluded from the study if they met the following exclusion criterion: had images with significant motion artifacts.

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committee of Union Hospital, Tongji Medical College, Huazhong University of Science and Technology (No. UHCT-IEC-SOP-016-03-01), and the requirement of individual consent for this retrospective analysis was waived.

CT scanning protocol

This study used a Philips 64-slice CT scanner (Incisive CT, Philips Healthcare, Amsterdam, The Netherlands). All the patients underwent scanning in a supine position under the respiratory voice command mode, with the scanning range extending from the lung apex to the lung base. The specific scanning parameters for the experimental group (Groups A and B: low-dose scanning protocols A1–B2; see Table 1) were as follows: tube voltage: 100 kVp with automatic tube current modulation; rotation speed: 0.5 r/s; dose right index (DRI): 4; collimation: 0.625×64; matrix: 512×512; and pitch: 0.8. The PI setting was adjusted to a high-intensity level (smooth), and the PI reconstruction modes were selected for the lung (lung window) and soft tissue (mediastinal window), with a slice thickness of 1 mm and a slice interval of 0.5 mm. After scanning, the thin-section lung and mediastinal window images were reconstructed using PI. Additionally, the low-dose raw data were processed with iDose iterative reconstruction (iDose4, level 4) to generate images with the same thickness, interval, and coverage as the PI-mode images, ensuring consistency. The filter functions used were “B” for the mediastinal window and “UB” for the lung window.

The specific scanning parameters for the control group (Group C: conventional-dose scanning protocols C1–C2; see Table 1) were as follows: tube voltage: 120 kVp with automatic tube current modulation; rotation speed: 0.5 r/s; DRI: 19; collimation: 0.625×64; matrix: 512×512; and pitch: 0.8. After scanning, the thin-section lung and mediastinal window images were reconstructed using iDose iterative reconstruction (iDose4, level 4), with the same filter functions applied (i.e., “B” for the mediastinal window and “UB” for the lung window).

For both groups, the window width and level for lung and mediastinal windows were set at [1,500, −500] and [200, 50], respectively. Upon completion of the scan, dose reports were automatically generated and recorded. Details of the scanning and reconstruction parameters, as well as group allocation, are provided in Table 1.

Principle of the PI reconstruction algorithm

The PI algorithm (21) is a novel AI-DLR algorithm developed by Philips Healthcare. It employs a supervised learning process to train the convolutional neural network (CNN) in a specific manner. Once trained, the CNN can reconstruct low-dose CT scan data to resemble images obtained from conventional-dose scans, including noise amplitude and noise texture. The training process of PI (21) is as follows: (I) clinical scan data from conventional-dose imaging are collected; (II) a sophisticated low-dose simulation technique is applied to generate low-dose scan data from conventional-dose data, accurately simulating photon and electronic noise in low-dose scans; (III) the conventional-dose scan data are reconstructed using the traditional FBP technique; and (IV) a CNN is trained to reproduce the appearance of conventional-dose FBP images using low-dose scan data (22). The CNN is trained with paired data of low-noise images (defined as the ground truth) and noisy images (23). Through this training process, parameter selection and node weights are optimized to maximize the consistency and accuracy of DLR. As a result, the PI algorithm reduces noise intensity while suppressing artifacts during CT image reconstruction. Unlike conventional iterative reconstruction methods, PI is trained to replicate the noise texture of FBP while significantly reducing noise (20,24,25).

Image quality evaluation

Objective evaluation

All the images from each group were imported into the graphical post-processing workstation (Philips IntelliSpace Portal 10.0, Philips Healthcare). A radiologist with 5 years of experience delineated regions of interest (ROIs) in six corresponding anatomical structures across all six sets of reconstructed images, including subcutaneous fat of the chest wall, pulmonary artery trunk, lung parenchyma, aortic trunk, cancellous bone of the vertebral body, and the erector spinae muscle. During the delineation process, the copy-and-paste function was used to ensure that ROIs of identical size were placed in corresponding regions across all image sets. Areas with homogeneous density were prioritized, while regions with obvious aeration or severe calcification were avoided. Measurements were obtained by selecting three consecutive slices and calculating their average value as the final result.

The CT value and standard deviation (SD) of each ROI were recorded, with the SD of the erector spinae muscle serving as the background noise level of the image. Further, the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were calculated, with the absolute values taken as the final results. The calculation formulas are expressed as follows:

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

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

Subjective evaluation

Two radiologists, with 5 and 10 years of experience, respectively, independently assessed the chest images of the experimental and control groups in a double-blind manner. They were blind to the reconstruction methods and dosage details for each group. In case of inconsistencies in their evaluations, consensus was reached through discussion. The grading criteria followed a five-point scale on which a score of 5 indicated that all chest tissues and structures were clearly displayed, with minimal image noise and no artifacts; a score of 4 indicated that the chest tissues and structures were relatively clear, with minor noise and small artifacts; a score of 3 indicated that the chest tissues and structures were visible, but with significant noise and a few artifacts; a score of 2 indicated that the chest tissues and structures were unclear, with substantial noise and noticeable artifacts; and a score of 1 indicated that the chest tissues and structures were unclear, with high noise levels and significant artifacts. Images with a score of ≥3 points were considered to meet the clinical diagnostic requirements.

Radiation dose

The dose report automatically generated by the equipment was used to record each patient’s volume CT dose index (CTDIvol) and dose-length product (DLP). The effective dose (ED) was calculated using the following formula: ED=DLP×k, where k is the tissue-weighting factor, with k for the chest being 0.014 (26).

Statistical analysis

The statistical analysis was performed using SPSS version 22.0 (IBM Corp., Armonk, NY, USA). For measurement data following a normal distribution, paired-sample t-tests were used for within-group comparisons, while independent-sample t-tests were used for between-group comparisons. For measurement data not following a normal distribution, Wilcoxon signed-rank tests were used for between-group comparisons. The consistency of ratings among the image analysts was assessed using the kappa statistic. A kappa value greater than 0.75 indicated excellent agreement, values between 0.4 and 0.75 indicated moderate to substantial agreement, and values ≤0.4 indicated poor agreement. A P value >0.05 indicated no statistically significant difference, while a P value <0.01 indicated a statistically significant difference.

Results

Clinical data description

This study retrospectively analyzed the data of 100 patients undergoing routine health examinations, of whom 66 were male and 34 were female. The patients had an average age of 49±13 years and an average body mass index (BMI) of 24.5±3.3 kg/m² (for details, see Table 2).

In the experimental and control groups, the mean CTDIvol values for the patients were 1.37±0.22 and 7.07±1.70 mGy, the mean DLP values were 56.36±9.82 and 296.8±80.72 mGy·cm, and the mean ED values were 0.79±0.14 and 4.16±1.13 mSv, respectively. Compared to the control group, the experimental group showed a reduction of 81.00% in the CTDIvol, DLP, and ED (for details, see Figure 1A-1C).

Figure 1 Objective evaluation, subjective scoring, and dose comparison of images in the experimental and control groups. (A-D) Compare the dose and subjective scoring between the experimental and control groups. (E-H) The bar charts of the CT, SD, SNR, and CNR values of the lung tissue ROIs in the experimental and control groups. A1 represents the mediastinal window images of the experimental group with low-dose PI reconstruction; A2 represents the lung window images of the experimental group with low-dose PI reconstruction; B1 represents the mediastinal window images of the experimental group with low-dose iDose4 reconstruction; B2 represents the lung window images of the experimental group with low-dose iDose4 reconstruction; C1 represents the mediastinal window images of the control group with standard-dose iDose4 reconstruction; C2 represents the lung window images of the control group with standard-dose iDose4 reconstruction. “ns” indicates no significant difference between groups (P>0.05); “****” indicates a highly significant difference of P<0.001. CNR, contrast-to-noise ratio; CT, computed tomography; CTDIvol, volume CT dose index; DLP, dose-length product; ED, effective dose; HU, Hounsfield unit; PI, precise imaging; ROI, region of interest; SD, standard deviation; SNR, signal-to-noise ratio.

Objective quality evaluation

Through the measurement, analysis, and calculation of data from the ROIs, the CT value, SD, SNR, and CNR of the six groups of images in the experimental groups (A1, A2) and (B1, B2), and the control group (C1, C2) followed a normal distribution. The results showed no significant differences in the CT values between the experimental groups (A1, A2) and (B1, B2) and the control group (C1, C2) across all ROIs (P>0.05). Similarly, no significant statistical differences were found in the SD, SNR, and CNR between the experimental group (A1, A2) and control group (C1, C2) (P>0.05), except for the SD values of subcutaneous fat in the chest wall and the erector spinae muscle between A1 and C1.

The SD values of the experimental group (A1, A2) were lower than those of the experimental group (B1, B2), while their SNR and CNR values were higher, and the differences were statistically significant (P<0.01), except for the CNR values of the pulmonary trunk and aortic arch. Additionally, the SD values of the experimental group (B1, B2) were higher than those of the control group (C1, C2), while their SNR and CNR values were lower, and the differences were statistically significant (P<0.01), except for the CNR values of the pulmonary trunk and aortic arch. Detailed results are shown in Figure 1 and Table 3.

Subjective quality evaluation

Two experienced radiologists subjectively evaluated the images of 100 patients in both the experimental and control groups. Inconsistencies arose in the scores of 12 cases, which were resolved through discussion and reassessment, ultimately achieving consensus. The kappa values for the ratings of the two radiologists for the experimental groups (A1, A2) and (B1, B2), as well as the control group (C1, C2), were 0.75, 0.78, and 0.81, respectively. The image quality and details of each group are illustrated in Figures 2,3. The subjective scores for the experimental groups (A1, A2) and (B1, B2), and the control group (C1, C2) were 4.40±0.53, 1.42±0.50, and 4.30±0.51, respectively. The image quality scores of the experimental group (A1, A2) and the control group (C1, C2) were significantly higher than those of the experimental group (B1, B2), while no statistically significant difference was observed between the scores of the experimental group (A1, A2) and the control group (C1, C2) (P>0.05). However, a statistically significant difference was found between the scores of the experimental group (B1, B2) and the control group (C1, C2) (P<0.01). The detailed scoring results are shown in Figure 1D.

Figure 2 Image comparison between the experimental and control groups. (A-C) Lung window images, while (D-F) are mediastinal window images. (A,B,D,E) The CT images of the level of the pulmonary trunk and aortic arch intersection for a 35-year-old male patient with a BMI of 24.36 kg/cm². (A,D) The low-dose images reconstructed using the PI high-level algorithm for the lung window and mediastinal window, respectively. (B,E) The low-dose images reconstructed using iDose4 (level 4) for the lung window and mediastinal window, respectively. (C,F) The CT images of the level of the pulmonary trunk and aortic arch intersection for a 31-year-old male patient with a BMI of 25.24 kg/cm², with (C,F) representing the conventional-dose images reconstructed using iDose4 (level 4) for the lung window and mediastinal window, respectively. The image scores for (A,D), (B,E), and (C,F) are 5, 1, and 5, respectively. BMI, body mass index; CT, computed tomography; PI, precise imaging.

Figure 3 Comparison of images reconstructed with PI (smooth level) and conventional iDose iterative reconstruction (iDose4, level 4) in the experimental group. The figures show the chest CT of a 35-year-old male patient with a BMI of 23.24 kg/cm². (A-D) The lung window and mediastinal window images reconstructed using low-dose combined with PI (at the smooth level). The chest tissues and structures are clearly depicted, with minimal image noise and no artifacts. The scattered solid pulmonary nodules are well-defined, with clear internal details, and the image score is 5. (E-H) The lung window and mediastinal window images reconstructed using low-dose combined with conventional iDose iterative reconstruction (iDose4, level 4). The chest tissues and structures are less clearly depicted, with higher image noise. The contours of the scattered solid pulmonary nodules are poorly defined, and internal details are blurred, with an image score of 2. BMI, body mass index; CT, computed tomography; PI, precise imaging.

Discussion

This study aimed to compare low-dose chest CT scanning combined with the deep learning-based AI-PI reconstruction algorithm against conventional-dose scanning combined with iDose iterative reconstruction (iDose4, level 4) to evaluate differences in image quality and radiation dose. In our study, the low-dose groups (A and B) used 100 kVp tube voltage, while the conventional-dose group used 120 kVp tube voltage. Lowering the tube voltage reduces the radiation dose and enhances soft-tissue contrast, aiding in the visualization of pulmonary nodules and mediastinal structures (27). We kept the tube current in the experimental group extremely low. Research shows that using 100 kVp tube voltage with iterative reconstruction technology can maintain good image details while reducing noise, ensuring acceptable image quality (28). A recent study also indicated that 100 kVp tube voltage is suitable for patients of average build (29). For larger patients, while a slight increase in voltage may sometimes be needed, 100 kVp can still serve as the initial choice and be adjusted as necessary.

The PI algorithm offers two methods for chest imaging: the soft-tissue algorithm, and the lung algorithm. Each method has five levels: sharper, sharp, standard, smooth, and smoother. In this study, the smooth level was chosen for PI reconstruction. Research by Greffier et al. (24) showed that as AI-DLR levels change from sharper to smoother, the image noise magnitude and optimum values change in different ways. The smooth level can effectively reduce noise while keeping enough image details. Additionally, at low doses (e.g., 0.4 mGy), the image smoothness of the smooth level is above average. Compared to the routine clinical iDose4 level 4, the smooth level has lower noise magnitude in two reconstruction kernels, resulting in clearer image quality. This balance allows the smooth level to deliver acceptable image quality across different doses. It is especially useful in clinical settings where both noise control and image detail are important.

In assessing chest image quality, ROIs, including muscle tissue, subcutaneous fat of the chest wall, lung tissue, the aortic trunk, the pulmonary trunk, and vertebrae, were selected for measurement. The results showed no statistically significant differences in the CT values among the experimental groups (A1, A2) and (B1, B2) and the control group (C1, C2) across all the ROIs (P>0.05). The CT value is not an absolute value but a relative value, and is minimally affected by different reconstruction algorithms and scanning conditions (e.g., tube current). These findings are consistent with the studies by Tomasi et al. (30) and van Stiphout et al. (31).

In terms of image noise and contrast, the SD values of the ROIs in the experimental group (A1, A2) and the control group (C1, C2) were lower than those of the experimental group (B1, B2), while the SNR and CNR values were higher, and the differences were statistically significant (P<0.01). Notably, this trend did not apply to the CNR values of the pulmonary trunk and aortic trunk. This indicates that under low-dose scanning conditions, images reconstructed by conventional iDose iterative reconstruction (iDose4, level 4) show higher noise levels compared to those reconstructed with the PI algorithm at low dose, as well as compared to those reconstructed by iDose iterative reconstruction (iDose4, level 4) at conventional-dose levels. Meanwhile, under low-dose scanning conditions, images reconstructed by conventional iDose iterative reconstruction (iDose4, level 4) exhibit lower SNR and CNR values compared to those reconstructed with the PI algorithm under low-dose conditions and conventional-dose iDose iterative reconstruction (iDose4, level 4). These findings are consistent with previous studies on deep learning–based AI reconstruction algorithms (32-34).

Additionally, no significant differences were observed in the SD, SNR, and CNR values between the experimental group (A1, A2) and control group (C1, C2) (P>0.05), while the experimental group (B1, B2) exhibited significant differences in these parameters compared to the control group (C1, C2) (P<0.01). This suggests that under low-dose conditions, the deep learning–based PI reconstruction algorithm, due to its superior noise-reduction capability, produces image quality comparable to that of iDose iterative reconstruction (iDose4, level 4) at conventional-dose levels. This conclusion aligns with previous research findings (35).

During specific measurements, slight differences were observed in the SD values of subcutaneous chest wall fat and the erector spinae muscle between A1 and C1 (P<0.05). This was likely due to a measurement bias resulting from ROI area instability. However, this discrepancy had a negligible effect on the overall objective assessment. Additionally, the CNR value of the pulmonary ROI in A1 was slightly lower than that in C1. This likely resulted from our dramatic reduction in tube current (DRI from 19 to 4) while lowering the tube voltage. Although this caused a significant noise increase, the SNR and CNR in A1 remained comparable to those in C1 (P>0.05), due to the excellent noise-reduction capability of the PI reconstruction algorithm. Detailed results are shown in Figure 1 and Table 3.

The scoring results of the two radiologists showed that the subjective evaluation scores for the experimental group (A1, A2) and the control group (C1, C2) were (4.40±0.53) and (4.30±0.51), respectively, while the score for the experimental group (B1, B2) was significantly lower at (1.42±0.50). Based on the five-point evaluation scale, the images reconstructed using a low-dose combined with PI reconstruction group (A1, A2), and those reconstructed using a conventional dose combined with iDose iterative reconstruction group (C1, C2) demonstrated clear visualization of thoracic structures with low image noise and minimal artifacts, meeting the requirements for clinical diagnosis. Conversely, the images from the low-dose combined with iDose iterative reconstruction group (B1, B2) exhibited blurred anatomical structures, high noise levels, and pronounced artifacts, rendering them unsuitable for clinical diagnosis. These findings indicate that a low dose combined with PI reconstruction can provide high-quality images comparable to those produced by a conventional dose with iDose iterative reconstruction, making it a viable option for clinical use. However, images reconstructed using a low dose combined with conventional iDose iterative reconstruction suffered from excessive noise and poor detail visibility, making them inadequate for clinical diagnosis. This finding further validates the superiority of deep learning-based AI (PI) algorithms in handling noise and preserving structural details at low-dose levels (36-39).

For radiation protection in health screening populations, the ideal scenario is to perform scans at low-dose levels without compromising image quality or diagnostic accuracy. In this study, the experimental and control groups had mean CTDIvol values of 1.37±0.22 and 7.07±1.70 mGy, mean DLP values of 56.36±9.82 and 296.80±80.72 mGy·cm, and mean ED values of 0.79±0.14 and 4.16±1.13 mSv, respectively. Compared with the control group, the experimental group achieved an 81.00% reduction in dose. This indicates that at an extremely low scanning dose, images processed using the PI reconstruction algorithm not only maintained image quality but also significantly reduced radiation exposure (25).

The ED, which is calculated by multiplying the DLP by International Commission on Radiological Protection (ICRP)-assigned tissue-weighting factors based on radiation cancer risk, offers a more comprehensive reflection of patients’ radiation dose and is crucial for assessing the radiation-related cancer risk (26). Further, this study found that the images obtained using a high-intensity PI algorithm combined with a low-dose scanning protocol (100 kVp, DRI =4) were comparable in quality to those obtained with iDose iterative reconstruction (iDose4, level 4) at a conventional dose (120 kVp, DRI =19). Meanwhile, the radiation dose for the chest CT scans was reduced to one-fifth of the conventional dose (approximately 0.8 mSv). These findings provide valuable insights for optimizing radiation protection in health screening populations.

This study had a number of limitations. First, during the actual measurement of the ROIs, the uncertainty in the size and positioning of the ROIs for the subcutaneous fat of the chest wall and the erector spinae muscle resulted in measurement errors that could not be completely avoided. This led to a slight statistical difference in the SD between the experimental group (A1, A2) and the control group (C1, C2) (P<0.05). Second, this study used the CT dose index and DLP as machine-output metrics, and the E-weighted index k was model-derived for radiation dose comparison, without considering patient size-related actual radiation dose differences. Third, this study was based solely on chest CT scans of health-checkup patients and did not include contrast-enhanced scans. Therefore, the performance of the PI algorithm in enhanced scans or vascular imaging was not evaluated; future research should explore its potential applications in these areas. Fourth, the sample in this study was limited to normal health-checkup patients, with a primary focus on the chest image quality and radiation dose in healthy individuals and no discussion on disease-specific aspects was included. Future studies could apply the PI reconstruction algorithm to research focused on specific diseases to assess its practical value in clinical diagnosis. Finally, the low-dose scanning protocol used in this study (DRI =4) was derived from a comparison with the standard-dose non-contrast chest CT scan of Philips Incisive CT (DRI =19), representing only one-fifth of the standard dose. However, previous studies have shown that there are various ultra-low dose scanning protocols available (40), and the extent to which the dose can be reduced further remains to be explored and validated in future research.

Conclusions

Images obtained using the low-dose scanning protocol combined with the PI reconstruction algorithm (at the smooth level) showed no significant differences to images obtained using conventional-dose iDose iterative reconstruction (iDose4, level 4) in terms of the objective evaluation metrics and subjective assessments (P>0.05). However, at low-dose levels, images obtained using conventional iDose iterative reconstruction (iDose4, level 4) showed significant differences to images obtained using conventional-dose iDose4 reconstruction in terms of both the objective data and subjective evaluations (P<0.01). This finding suggests that the deep learning–based AI-PI reconstruction algorithm (at the smooth level) can significantly reduce the radiation dose while maintaining high-quality images in chest CT scans. Therefore, this approach should be considered a preferred choice for health-checkup patients.

Acknowledgments

We appreciate the contributions of the colleagues who participated in data analysis, manuscript drafting, and critical revisions.

Footnote

Reporting Checklist: The authors have completed the STROBE reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-685/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-685/dss

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (https://qims.amegroups.com/article/view/10.21037/qims-2025-685/coif). S.G. is employed by Philips Healthcare; the company provided no financial or material support for this work. 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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committee of Union Hospital, Tongji Medical College, Huazhong University of Science and Technology (No. UHCT-IEC-SOP-016-03-01) and individual consent for this retrospective analysis was waived.

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. Mohammadinejad P, Mileto A, Yu L, Leng S, Guimaraes LS, Missert AD, Jensen CT, Gong H, McCollough CH, Fletcher JG CT. Noise-Reduction Methods for Lower-Dose Scanning: Strengths and Weaknesses of Iterative Reconstruction Algorithms and New Techniques. Radiographics 2021;41:1493-508. [Crossref] [PubMed]
2. Mileto A, Guimaraes LS, McCollough CH, Fletcher JG, Yu L. State of the Art in Abdominal CT: The Limits of Iterative Reconstruction Algorithms. Radiology 2019;293:491-503. [Crossref] [PubMed]
3. Hara AK, Paden RG, Silva AC, Kujak JL, Lawder HJ, Pavlicek W. Iterative reconstruction technique for reducing body radiation dose at CT: feasibility study. AJR Am J Roentgenol 2009;193:764-71. [Crossref] [PubMed]
4. Li J, Gao J, Wang G, Cui D, Deng K. A phantom study using dual-energy spectral computed tomography imaging: Comparison of image quality between two adaptive statistical iterative reconstruction (ASiR, ASiR-V) algorithms for evaluating ground-glass nodules of the lung. J Cancer Res Ther 2021;17:1742-7. [Crossref] [PubMed]
5. Schmidt-Holtz J, Laqmani A, Butscheidt S, Kurfürst M, Avanesov M, Behzadi C, Spink C, Veldhoen S, Nagel HD, Adam G, Regier M. Iterative Model Reconstruction (IMR) in MDCT Below 2 mSv for the Detection of Urinary Calculi: Diagnostic Accuracy and Image Quality in Comparison to Filtered Back-Projection and 4th Generation Iterative Reconstruction (iDose4). Rofo 2018;190:630-6. [Crossref] [PubMed]
6. 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]
7. Euler A, Solomon J, Marin D, Nelson RC, Samei E. 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]
8. 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]
9. Fair E, Profio M, Kulkarni N, Laviolette PS, Barnes B, Bobholz S, Levenhagen M, Ausman R, Griffin MO, Duvnjak P, Zorn AP, Foley WD. Image Quality Evaluation in Dual-Energy CT of the Chest, Abdomen, and Pelvis in Obese Patients With Deep Learning Image Reconstruction. J Comput Assist Tomogr 2022;46:604-11. [Crossref] [PubMed]
10. Heinrich A, Yücel S, Böttcher B, Öner A, Manzke M, Klemenz AC, Weber MA, Meinel FG. Improved image quality in transcatheter aortic valve implantation planning CT using deep learning-based image reconstruction. Quant Imaging Med Surg 2023;13:970-81. [Crossref] [PubMed]
11. Hsieh J, Liu E, Nett B, Tang J, Thibault JB, Sahney S. A new era of image reconstruction: TrueFidelity. Technical white paper on deep learning image reconstruction. 2019. Available online: https://www.gehealthcare.com/-/jssmedia/040dd213fa89463287155151fdb01922.pdf?srsltid=AfmBOor_Z27xrQSoEKJ9V2cJgAqcRZ2DE4rH4fQRSzuIG8QKdbGT6CPR
12. Boedeker K. AiCE Deep Learning Reconstruction: Bringing the power of Ultra-High Resolution CT to routine imaging. 2019. Available online: https://global.medical.canon/publication/ct/2019WP_AiCE_Deep_Learning
13. Akagi M, Nakamura Y, Higaki T, Narita K, Honda Y, Zhou J, Yu Z, Akino N, Awai K. Deep learning reconstruction improves image quality of abdominal ultra-high-resolution CT. Eur Radiol 2019;29:6163-71. [Crossref] [PubMed]
14. Akagi M, Nakamura Y, Higaki T, Narita K, Honda Y, Awai K. Deep learning reconstruction of equilibrium phase CT images in obese patients. Eur J Radiol 2020;133:109349. [Crossref] [PubMed]
15. Franck C, Zhang G, Deak P, Zanca F. 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]
16. Nakamoto A, Onishi H, Ota T, Honda T, Tsuboyama T, Fukui H, Kiso K, Matsumoto S, Kaketaka K, Tanigaki T, Terashima K, Enchi Y, Kawabata S, Nakasone S, Tatsumi M, Tomiyama N. 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]
17. Wang H, Yue S, Liu N, Chen Y, Zhan P, Liu X, Shang B, Wang L, Li Z, Gao J, Lyu P. Deep learning reconstruction vs standard reconstruction for abdominal CT: the influence of BMI. Eur Radiol 2024;34:1614-23. [Crossref] [PubMed]
18. Kazimierczak W, Kazimierczak N, Wilamowska J, Wojtowicz O, Nowak E, Serafin Z. Enhanced visualization in endoleak detection through iterative and AI-noise optimized spectral reconstructions. Sci Rep 2024;14:3845. [Crossref] [PubMed]
19. Huang X, Zhao W, Wang G, Wang Y, Li J, Li Y, Zeng Q, Guo J. Improving image quality with deep learning image reconstruction in double-low-dose head CT angiography compared with standard dose and adaptive statistical iterative reconstruction. Br J Radiol 2023;96:20220625. [Crossref] [PubMed]
20. Greffier J, Frandon J, Durand Q, Kammoun T, Loisy M, Beregi JP, Dabli D. Contribution of an artificial intelligence deep-learning reconstruction algorithm for dose optimization in lumbar spine CT examination: A phantom study. Diagn Interv Imaging 2023;104:76-83. [Crossref] [PubMed]
21. Philips. AI for significantly lower dose and improved image quality. Precise Image. Available online: https://www.philips.com/c-dam/b2bhc/master/resource-catalog/landing/precise-suite/incisive_precise_image.pdf
22. Greffier J, Durand Q, Frandon J, Si-Mohamed S, Loisy M, de Oliveira F, Beregi JP, Dabli D. Improved image quality and dose reduction in abdominal CT with deep-learning reconstruction algorithm: a phantom study. Eur Radiol 2023;33:699-710. [Crossref] [PubMed]
23. Brady SL. Implementation of AI image reconstruction in CT-how is it validated and what dose reductions can be achieved. Br J Radiol 2023;96:20220915. [Crossref] [PubMed]
24. Greffier J, Si-Mohamed S, Frandon J, Loisy M, de Oliveira F, Beregi JP, Dabli D. 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]
25. Barca P, Domenichelli S, Golfieri R, Pierotti L, Spagnoli L, Tomasi S, Strigari L. 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]
26. Deak PD, Smal Y, Kalender WA. Multisection CT protocols: sex- and age-specific conversion factors used to determine effective dose from dose-length product. Radiology 2010;257:158-66. [Crossref] [PubMed]
27. Vonder M, Dorrius MD, Vliegenthart R. Latest CT technologies in lung cancer screening: protocols and radiation dose reduction. Transl Lung Cancer Res 2021;10:1154-64. [Crossref] [PubMed]
28. Petritsch B, Kosmala A, Weng AM, Bley TA. Tin-filtered 100 kV ultra-low-dose CT of the paranasal sinus: Initial clinical results. PLoS One 2019;14:e0216295. [Crossref] [PubMed]
29. Yao Y, Su X, Deng L, Zhang J, Xu Z, Li J, Li X. Effects of tube voltage, radiation dose and adaptive statistical iterative reconstruction strength level on the detection and characterization of pulmonary nodules in ultra-low-dose chest CT. Cancer Imaging 2024;24:123. [Crossref] [PubMed]
30. Tomasi S, Szilagyi KE, Barca P, Bisello F, Spagnoli L, Domenichelli S, Strigari L. A CT deep learning reconstruction algorithm: Image quality evaluation for brain protocol at decreasing dose indexes in comparison with FBP and statistical iterative reconstruction algorithms. Phys Med 2024;119:103319. [Crossref] [PubMed]
31. van Stiphout JA, Driessen J, Koetzier LR, Ruules LB, Willemink MJ, Heemskerk JWT, van der Molen AJ. The effect of deep learning reconstruction on abdominal CT densitometry and image quality: a systematic review and meta-analysis. Eur Radiol 2022;32:2921-9. [Crossref] [PubMed]
32. Jiang B, Li N, Shi X, Zhang S, Li J, de Bock GH, Vliegenthart R, Xie X. Deep Learning Reconstruction Shows Better Lung Nodule Detection for Ultra-Low-Dose Chest CT. Radiology 2022;303:202-12. [Crossref] [PubMed]
33. 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]
34. Ye K, Xu L, Pan B, Li J, Li M, Yuan H, Gong NJ. Deep learning-based image domain reconstruction enhances image quality and pulmonary nodule detection in ultralow-dose CT with adaptive statistical iterative reconstruction-V. Eur Radiol 2025;35:3739-49. [Crossref] [PubMed]
35. Zhang F, Liu J, Liu Y, Zhang X. Research progress of deep learning in low-dose ct image denoising. Radiat Prot Dosimetry 2023;199:337-46. [Crossref] [PubMed]
36. 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]
37. Son W, Kim M, Hwang JY, Kim YW, Park C, Choo KS, Kim TU, Jang JY. Comparison of a Deep Learning-Based Reconstruction Algorithm with Filtered Back Projection and Iterative Reconstruction Algorithms for Pediatric Abdominopelvic CT. Korean J Radiol 2022;23:752-62. [Crossref] [PubMed]
38. Yunaga H, Miyoshi H, Ochiai R, Gonda T, Sakoh T, Noma H, Fujii S. Image Quality and Lesion Detection of Multiplanar Reconstruction Images Using Deep Learning: Comparison with Hybrid Iterative Reconstruction. Yonago Acta Med 2024;67:100-7. [Crossref] [PubMed]
39. Xia Z, Liu J, Kang Y, Wang Y, Hu D, Zhang Y. Dynamic controllable residual generative adversarial network for low-dose computed tomography imaging. Quant Imaging Med Surg 2023;13:5271-93. [Crossref] [PubMed]
40. Beregi JP, Greffier J. Low and ultra-low dose radiation in CT: Opportunities and limitations. Diagn Interv Imaging 2019;100:63-4. [Crossref] [PubMed]