Comparison of DLIR-H and ASIR algorithms for image reconstruction in low-dose chest CT of pediatric mycoplasma pneumoniae pneumonia: a cross-sectional study

Introduction

Mycoplasma pneumoniae pneumonia (MPP) is a major pathogen of community-acquired pneumonia (CAP) in children globally, and accounts for numbers of pediatric respiratory hospitalizations (1-3). Given the non-specific clinical manifestations of pediatric MPP, chest computed tomography (CT) has become an indispensable imaging modality for early lesion detection, severity stratification, and prognostic evaluation, because of its superior spatial resolution and sensitivity to subtle parenchymal changes compared with chest radiography (1,4,5). However, its application as a screening modality for pediatric patients has been hampered by public concerns about the cumulative burden of radiation exposure (6). Therefore, it is critical to adopt low-dose chest CT protocols in pediatric practice; however, aggressive dose reduction inevitably compromises image quality due to increased quantum noise, potentially obscuring subtle MPP lesions.

To address the trade-off between radiation dose reduction and diagnostic image quality, CT image reconstruction technology has undergone three pivotal generations of evolution: filtered back projection (FBP), hybrid iterative reconstruction (HIR), and deep learning image reconstruction (DLIR). FBP is characterized by high computational efficiency and straightforward implementation. However, it inherently propagates quantum noise from raw projection data to the final image, resulting in considerable noise in the images at low radiation doses and making it unsuitable for pediatric applications (7). HIR represents a second-generation technological advancement, that integrates statistical iterative methods with FBP to suppress noise while preserving anatomical details. Adaptive statistical iterative reconstruction (ASIR), a widely used representative HIR algorithm in pediatric imaging, reduces noise by blending iterative and FBP kernels at adjustable weighting factors (8). However, HIR has certain limitations. Specifically, excessive iterative weighting will lead to noise texture degradation and the generation of “waxy” or smooth artifacts, which may obscure fine anatomical structures (e.g., lobular septa) and subtle multiple alveolar lesions [e.g., ground-glass opacity (GGO)] (9). Additionally, HIR fails to fully separate noise from low-contrast signals, resulting in compromised contrast resolution, which is critical for detecting mild parenchymal changes of early MPP. These limitations restrict the maximum achievable dose reduction with HIR and create an unmet clinical need for a more robust reconstruction approach in pediatric low-dose CT. With respect to DLIR, algorithms developed by General Electric (GE) Healthcare feature a deep neural network using low doses, which is trained with high-quality FBP datasets (TrueFidelity™, GE Healthcare, Chicago, IL, USA) to learn how to achieve radiation dose reduction and reconstruct CT images without changing the noise texture or affecting the anatomy (10). Unlike HIR, which relies on mathematical iterative optimization, DLIR directly learns to suppress noise while retaining natural tissue texture and low-contrast details, fundamentally overcoming the limitation of HIR regarding the generation of “waxy” artifacts. Among the DLIR variants, high-strength DLIR (DLIR-H) is optimized for maximum noise reduction and contrast preservation, making it theoretically well-suited for the extreme dose constraints of pediatric imaging. While DLIR may significantly improve the image quality of coronary angiography (10), thoracic low-dose CT (11), abdominal low-dose CT (12), and cerebral low-dose CT (13) in adults, its clinical efficacy in pediatric MPP remains largely unexplored.

Pediatric MPP presents unique challenges for low-dose CT reconstruction that further underscore the need for focusing on DLIR-H. First, MPP lesions are often subtle, consisting of patchy GGOs and mild consolidation, which are highly susceptible to blurring by HIR which tends to smoothen artifacts. Second, compared with adults, the pediatric thorax has smaller anatomical structures and lower tissue contrast, which amplifies the impact of reconstruction-related resolution loss on diagnostic accuracy. Third, the objective of minimizing cumulative radiation exposure in children necessitates the development of a reconstruction algorithm that can enable further dose reduction beyond the limits of HIR without sacrificing diagnostic quality.

The purpose of this cross-sectional study was to compare the performance of DLIR with that of ASIR in low-dose CT for pediatric MPP. This study sought to validate whether DLIR-H can overcome the limitations of HIR in this specific clinical context. The scientific innovation of this study lies in achieving an optimal balance between radiation safety and high lesion visualization, filling the gap in DLIR-H application research for the imaging-based diagnosis of pediatric MPP. The results of this study are expected to provide critical evidence for optimizing CT reconstruction strategies in pediatric MPP. We present this article in accordance with the STROBE reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-1-923/rc).

Methods

Study setting and participant recruitment

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This research was conducted at the First Teaching Hospital of Tianjin University of Traditional Chinese Medicine, and was approved by the hospital’s internal review board (No. TYLL2024[Z] 025); the informed consent was waived because of the retrospective design. Consecutive pediatric patients from September 2023 to March 2024 were enrolled retrospectively. The specific study setting was shown in Appendix 1.

All pediatric pneumonia cases underwent initial screening via Electronic Medical Record System (EMRs) and Picture Archiving and Communication System (PACS). Two independent researchers conducted inclusion and exclusion validation, followed by standardized data collection for enrollment. The inclusion criteria were as follows: (I) <18 years old and first admitted for pneumonia; (II) first chest CT performed during hospitalization; (III) positive for MP-IgM antibody and negative for seven-item respiratory pathogen antibodies; and (IV) complete raw CT data available for multi-algorithm reconstruction. The exclusion criteria were as follows: (I) incomplete clinical, laboratory, or imaging data; (II) poor cooperation resulting in unqualified CT image quality; (III) a body mass index (BMI) ≥25 kg/m²; (IV) failed algorithm reconstruction; and (V) a history of other severe pulmonary diseases or systemic diseases affecting imaging findings. The flowchart of participant recruitment was detailed in Appendices 2,3.

The methodology overview of this study was shown in Figure 1.

Figure 1 Methodology overview and participant recruitment flowchart of the study. A schematic diagram illustrating the entire research workflow, including five core stages: (1) participant recruitment; (2) CT image acquisition; (3) CT image reconstruction; (4) radiation dose assessment; (5) image processing segmentation. Sample size estimation was performed using IBM SPSS Statistics 26.0 with α=0.05 and power=0.8, confirming sufficient statistical power for the enrolled cohort. ANOVA, analysis of variance; ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; DLIR, deep learning image reconstruction; ED, effective dose; SSDE, size-specific dose estimate.

Sample size estimation

The sample size calculation was performed via IBM SPSS Statistics 26.0 software, with the lesion contrast-to-noise ratio (CNR) of the MPP lesions as the primary outcome. The preset significance level (α) was 0.05 and the test power was 0.8. Effect size (one-way analysis of variance [ANOVA] f=0.43, Cohen’s d=0.33) and the intraclass correlation coefficient (ICC) reliability analysis were used to determine a minimum sample size of 60. The final cohort of 142 patients provided sufficient statistical power. Detailed steps for sample size calculation are provided in Appendix 4.

CT image acquisition and reconstruction

All patients underwent low-dose chest CT examination with a 256-slice volume CT scanner (Revolution Apex CT, GE Healthcare, Chicago, IL, USA). Standardized breath-hold training for the patients was performed before the scan. The specific CT scan procedures are detailed in Appendix 5, with “low-dose CT” defined per Chinese guidelines (Appendix 6). Raw data were reconstructed with hybrid iterative reconstruction (ASIR, GE Healthcare) at 0%, 50%, and 80% blending factors, and DLIR (TrueFidelity, GE Healthcare) at low-strength level (DLIR-L), medium-strength level (DLIR-M), and DLIR-H, using a standard kernel and a 1.25 mm slice thickness/interval.

Radiation dose assessment

Radiation dose parameters, including the volume CT dose index (CTDI_vol, mGy), dose-length product (DLP, mGy·cm), and effective dose (ED, mSv) of radiation, were recorded. ED = DLP × k, where k is the chest-specific radiation dose conversion coefficient (k=0.014 mSv/mGy·cm) recommended by the International Commission on Radiological Protection (ICRP) (14). Considering the possible impact of children’s body size on the radiation dose, this study also incorporated the size-specific dose estimate (SSDE) parameter per American Association of Physicists in Medicine (AAPM) Report 204 (Appendix 7). All the parameters were compared with the Chinese pediatric CT radiation dose expert consensus and guidelines.

Segmentation and reconstruction of CT raw image

Reconstructed CT images were segmented with a 3D Slicer image computing platform (https://www.slicer.org/) to strictly exclude non-pulmonary tissues, such as the mediastinum and thoracic wall. The segmented data were imported into the OpenCV database (version 3.7.9, Python Software Foundation, Wilmington, DE, USA) for subsequent quantitative analysis. To eliminate interference from blood vessels and bronchi, the Hounsfield unit (HU) range of the parenchyma was defined as −1,000 to −700 HU. The above procedure for standardized segmentation was uniformly applied to six groups of CT images reconstructed with different algorithms (Figure 2).

Figure 2 Representative low-dose chest CT images reconstructed and segmentation by six different algorithms in a pediatric MPP patients. ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; DLIR, deep learning image reconstruction; MPP, Mycoplasma pneumoniae pneumonia.

Quantitative image quality evaluation

All reconstructed CT images were transmitted to GE Advantage Workstation 4.7 (GE Healthcare, USA) for standardized quantitative image quality evaluation, which was completed by two radiologists with more than 15 years of clinical experience in chest imaging diagnosis. They were both blinded to the reconstruction algorithm of the evaluated images.

The normal lung parenchyma and the MPP lesion on three consecutive transverse slices were selected to place a region of interest (ROI) with a size of approximately 20 mm² (Figure 3 and Appendix 8). The CT value and standard deviation (SD) value were measured and the mean value was calculated as the final value. The SD value of the subcutaneous fat was used to represent the image noise, and was measured by using the same ROI placement method. The signal-to-noise ratio (SNR) of the lung parenchyma and CNR of the MPP lesion were calculated (Eqs. [1,2]).

SNRlungparenchyma=CTlungparenchymaSDlungparenchyma[1]

CNRMPP=CTMPP−CTlungparenchymaSDlungparenchyma[2]

Figure 3 ROI placement and measurement for quantitative image quality evaluation. Schematic of ROI positioning on transverse CT slices: (A) the normal lung parenchyma and the MPP lesion were selected to place a ROI with a size of about 20 mm², and the replication function was used to avoid manual error; (B) delineation of ROI in MPP lesions in enlarged image; (C) delineation of ROI in the normal lung parenchyma in enlarged image; (D) the SD value of subcutaneous fat was used to represent image noise; (E) delineation of ROI in the subcutaneous fat in enlarged image. CT, computed tomography; MPP, Mycoplasma pneumoniae pneumonia; ROI, region of interest; SD, standard deviation.

The results determined by the two radiologists were averaged to eliminate observer measurement bias (Figure 4). The clinical interpretation of the negative SNR value is provided in Appendix 9.

Figure 4 Quantitative and qualitative image quality evaluation. The quantitative evaluation included the following parameters: the SNR of the lung parenchyma, the CNR of the MPP lesion, and background noise. The qualitative evaluation was scored using a standardized 5-point scoring method. ASIR, adaptive statistical iterative reconstruction; CNR, contrast-to-noise ratio; CT, computed tomography; DLIR-H, deep learning image reconstruction at high-strength level; HU, Hounsfield unit; MPP, Mycoplasma pneumoniae pneumonia; ROI, region of interest; RSNA, Radiological Society of North America; SD, standard deviation; SNR, signal-to-noise ratio.

Qualitative analysis of image quality

Two other senior radiologists with 15 years of experience in chest imaging diagnosis performed a double-blind qualitative evaluation, which was divided into two standardized stages to ensure the reliability of the evaluation results: the training stage and the evaluation stage (Appendix 10). The image quality was scored using a standardized 5-point scoring method: score 1= non-diagnostic, score 2= sub-diagnostic, score 3= diagnostic, score 4= good, and score 5= excellent. The images of each patient were scored three times. The average value of the scores determined by the two radiologists’ was considered the final score. One week later, all the images were re-scored. Intra- and inter-observer consistency analyses were performed (Figure 4).

Statistical analysis

All the statistical analyses were performed via IBM SPSS Statistics 26.0.0.

The Kolmogorov-Smirnov test was used to test the normal distribution of continuous variables; data of the normally distributed variables are expressed as the mean ± SD while that of the non-normally distributed variables are presented as the medians [interquartile ranges (IQRs)]. Categorical variables are expressed as numbers (frequencies). Group comparisons were conducted using chi-squared test (categorical), the Kruskal-Wallis H test (non-normal continuous), and one-way ANOVA (normal continuous). Post-hoc pairwise comparisons with Bonferroni correction were subsequently performed to identify the specific statistically significant differences between individual algorithm groups. The ICC was used to quantify observer consistency (substantial agreement defined as an ICC >0.7). All statistical tests were two-sided, and a P value <0.05 was considered to indicate statistical significance. All the data generated or analyzed during the study are available from the corresponding author upon reasonable request.

Results

Population characteristics

Among the 235 patients, 142 were eligible for the final analysis, with a median age of 9 years (IQR, 6, 10 years) and a mean BMI of 16.15 kg/m² (15.05, 16.90 kg/m²). The cohort comprised 77 boys (54.2%) and 65 girls (45.8%). The specific reasons for the exclusion of 93 patients are detailed in Appendix 11.

Clinical symptoms and laboratory test results

All the enrolled patients presented with typical respiratory symptoms associated with MPP: fever (129 cases, mild: 18, moderate: 45, severe: 66); cough (131 cases, mild: 4, moderate: 19, severe: 108); and wheezing (5 cases, mild: 3, moderate: 2, severe: 0). Symptom severity was classified according to the clinical diagnostic criteria for pediatric respiratory infectious diseases, and no other severe concomitant symptoms were observed in the cohort.

All patients underwent routine peripheral blood tests and inflammatory marker detection before CT examination. The laboratory indices are nonnormally distributed and are expressed as the medians (IQRs): peripheral blood white blood cell count (WBC): 8.04×10⁹ (6.45×10⁹, 10.18×10⁹)/L; C-reactive protein (CRP): 14.61 (7.77, 35.64) mg/L; procalcitonin (PCT): 0.25 (0.07, 1.45) ng/mL; and lactate dehydrogenase (LDH): 281.85 (251.28, 345.73) U/L. The clinical and demographic characteristics are given in Table 1 and Table S1, and Figure S1.

Radiation dose evaluation

The radiation dose parameters are all non-normally distributed (Table S1) and are expressed as median (IQR): CTDI_vol: 2.97 (2.55, 7.37) mGy; DLP: 83.52 (59.79, 180.45) mGy·cm, ED: 1.17 (0.84, 2.53) mSv; and SSDE: 3.08 (2.54, 5.63) mGy. The specific distributions of the radiation dose parameters by patient weight are shown in Table 2 and Figure S2. All the radiation dose parameters were lower than the reference levels recommended in three Chinese pediatric CT radiation dose guidelines (14-17), confirming that the scanning protocol strictly complied with clinical low-dose CT requirements for pediatric patients. Notably, radiation dose parameters showed a mild positive correlation with patient weight (Table S2), but even in the highest weight stratification, all dose indices remained below the guideline reference values.

Objective evaluation

The SNR of the lung parenchyma, the CNR of the MPP lesion, and background noise were all nonnormally distributed (Table S1). Overall, the result of the Kruskal-Wallis H test revealed statistically significant differences in the three objective indicators among the six reconstruction algorithms (all P<0.001; Table 3). Post-hoc pairwise comparisons with a Bonferroni correction were conducted for all intergroup differences.

Background noise gradually decreased with an increase in the ASIR blending factor and DLIR strength. DLIR-H exhibited the lowest background noise at 21.69 (16.52, 29.41), which was approximately 38.3% lower than that of ASIR-80% [35.13 (24.50, 45.96)]. ASIR-80% exhibited the highest SNR of the lung parenchyma at −14.46 (−18.67, −10.92); DLIR-H exhibited a SNR of the lung parenchyma value of −21.96 (−30.60, −17.48) (the negative SNR was normal for lung tissue; Appendix 9). The CNR of the MPP lesion increased with increasing ASIR blending factor and DLIR strength. DLIR-H exhibited the highest CNR of the MPP lesion value [28.22 (21.62, 35.39)], 14.7% higher than the exhibited by ASIR-80% [24.60 (18.59, 33.91)] (Figure 5).

Figure 5 Quantitative comparison of objective image quality indicators among six reconstruction algorithms. Violin plots illustrating differences in four key objective metrics: (A) CNR of MPP lesion (a higher DLIR-H value indicates better lesion-to-parenchyma contrast); (B) SNR of lung parenchyma (negative values are normal due to low lung CT density relative to noise); (C) background noise (SD of subcutaneous fat, a lower value of DLIR-H indicates less noise); (D) subjective quality scores. Kruskal-Wallis H test revealed significant differences among groups (all P<0.001). DLIR-H exhibits the highest CNR (28.22), lowest background noise (21.69), and highest subjective score (4.83±0.400), while ASIR-80% shows the highest lung parenchyma SNR (−14.46). ASIR, adaptive statistical iterative reconstruction; CNR, contrast-to-noise ratio; CT, computed tomography; DLIR, deep learning image reconstruction; DLIR-H, DLIR at high-strength level; DLIR-L, DLIR at low-strength level; DLIR-M, DLIR at medium-strength level; MPP, Mycoplasma pneumoniae pneumonia; SD, standard deviation; SNR, signal-to-noise ratio.

Subjective evaluation

The results of the Kruskal-Wallis H test revealed significant differences in subjective image quality scores among the six reconstruction algorithms (H=246.35, P<0.001), the detailed scores are presented in Table 4. Post-hoc pairwise comparisons with Bonferroni correction confirmed that DLIR-H achieved the highest subjective score at 4.83±0.400, which was significantly higher than that of all the other algorithms (all P<0.001). The subjective scores of both algorithm groups demonstrated a stepwise improvement as the reconstruction strength/blending factor increased. Pairwise comparisons revealed that compared with DLIR-L (P=0.03) but not DLIR-M (P=0.29), ASIR-80% had a significantly higher subjective score (Table S3).

Results of the observer consistency analysis

Both the intra-observer and inter-observer consistency analyses of the subjective scores exhibited substantial agreement: the intra-observer ICCs were 0.89 [95% confidence interval (CI): 0.86–0.91, P<0.001] and 0.86 (95% CI: 0.82–0.89, P<0.001), and the inter-observer ICC was 0.83 (95% CI: 0.79–0.86, P<0.001). Details of the observer consistency analysis are provided in Appendix 12, which indicate the high reliability of the subjective evaluation results.

DLIR-H images demonstrated minimal background noise, the clearest display of subtle MPP lesions (e.g., GGOs), and the best natural tissue texture, which significantly improved the radiologists’ diagnostic confidence for MPP. ASIR-80% images demonstrated moderate noise reduction but presented mild “waxy” artifacts in partial lung parenchyma, which slightly obscured fine anatomical structures and subtle lesions. The DLIR-L and ASIR-0% images exhibited relatively obvious background noise, making it more difficult to identify mild MPP lesions. A typical case of subjective image-quality comparison among different reconstruction algorithms is shown in Figure 6, which visually demonstrates the differences in noise level, display of lesion detail, and tissue contrast among the six algorithms.

Figure 6 Typical case demonstrating the superiority of DLIR-H in visualizing MPP lesions. A female patient, aged 11 years, presented with recurrent fever for 8 days. Laboratory examinations revealed a positive result for Mycoplasma pneumoniae-DNA amplification qualitative test. The nucleic acid tests for influenza A virus, influenza B virus, respiratory syncytial virus-RNA amplification qualitative test, human rhinovirus-RNA amplification qualitative test, and adenovirus DNA amplification qualitative test were all negative. The CRP level was 22.300 mg/L (normal range, 0.000–10.000 mg/L), and the fibrinogen level was 4.95 g/L (normal range, 2.00–4.00 g/L). (A) X-ray examination indicated inflammation in the right lower lobe. (B-H) The CT images of this patient. (B) The mediastinal window. (C-H) The reconstructed images of the lung window using hybrid iterative reconstruction with blending factors of 0%, 50%, and 80%, and DLIR at low-strength, medium-strength, and high-strength levels, respectively. An irregular patchy soft-tissue density shadow was observed in the middle and lower lobes of the right lung, accompanied by air bronchogram and tree-in-bud sign, and the edge of the lesion was indistinct. Among the six images, the image reconstructed by DLIR-H (H) exhibited the lowest background noise and the highest SNR and CNR. Simultaneously, the subjective image quality scores assigned by two radiologists for DLIR-H were both 5 points. (I) A follow-up chest X-ray conducted 9 days later demonstrated that the inflammation in the right lower lung lobe had been more absorbed compared to the previous state. CNR, the contrast to noise ratio; CRP, C-reactive protein; CT, computed tomography; DLIR, deep learning image reconstruction; DLIR-H, deep learning image reconstruction at high-strength level; DNA, deoxyribonucleic acid; RNA, ribonucleic acid; SNR, the signal to noise ratio.

Discussion

In this retrospective cross-sectional study, we systematically compared the image quality of six reconstructions and algorithms in low-dose chest CT for pediatric MPP, with an emphasis on validating the clinical value of DLIR-H. In line with the STROBE guidelines for reporting observational studies, our findings comprehensively indicate that DLIR-H attained an optimal equilibrium between radiation safety and diagnostic image quality, thereby fulfilling the key unmet clinical requirements in pediatric thoracic imaging. There were three significant findings. First, among the six algorithms, DLIR-H demonstrated the lowest background noise and the highest CNR of the MPP lesion, which significantly improved the visualization of subtle MPP lesions. Second, the subjective image quality score corroborated the superiority of DLIR-H in preserving the natural tissue texture and reducing artifacts. Third, the low effective dose conforms to the Chinese pediatric CT radiation dose guidelines, thus minimizing the cumulative radiation risk in children.

With respect to pediatric respiratory tract infections, fever combined with severe cough can serve as a key clinical indicator for suspected MPP. The low incidence of wheezing suggests that its absence should not exclude an MPP diagnosis, especially in younger/lower-weight children. Clinicians should focus on fever and cough patterns rather than relying on wheezing as a diagnostic criterion.

When assessing disease severity in high-weight/older children, their elevated baseline levels of WBC, CRP, and PCT should be considered. The baseline values of inflammatory markers in such children significantly differ from those in low-weight/young children, and a uniform cutoff value cannot be directly applied to determine disease severity. Specifically, the clinical value of CRP >20 mg/L or PCT >0.2 ng/mL in high-weight children for indicating severe infection is significantly lower than in low-weight children. A comprehensive evaluation combining overall symptoms and other test results is required to avoid misdiagnosis based on a single indicator. For patients with significantly elevated inflammatory markers, as they often exhibit a stronger systemic inflammatory response, prompt anti-inflammatory treatment should be intensified to control progression. All MPP patients require close monitoring of LDH levels. Low-weight/young children have higher baseline LDH levels and are more prone to tissue damage, and dynamic monitoring helps in early detection of damage and timely intervention.

Notably, the SNR of the lung parenchyma of the DLIR variants was lower than that of the ASIR algorithms. As clarified in Appendix 9, this negative SNR is a normal consequence of the lung’s low CT value relative to the positive noise, and the absolute value remains clinically interpretable. More importantly, the DLIR-H’s high CNR of the MPP lesion value offsets this difference in the SNR, as it prioritizes the contrast between MPP lesions and normal lung tissue, which is considerably more relevant for clinical diagnosis than the global parenchymal SNR. These finding align with the clinical requirements for pediatric MPP imaging, where detecting subtle lesions (e.g., patchy GGO) is more important than general parenchymal noise metrics.

The fundamental distinction between DLIR and ASIR lies in their technical principles (7,11,18,19). DLIR-H leverages deep neural networks trained on high-quality datasets to directly distinguish signals from noise, preserving the natural tissue texture while maximizing noise suppression (20,21). Compared with ASIR-80%, DLIR-H achieved a background noise reduction of 38.3% and an increase in the CNR by 14.7%, thereby overcoming the limitations of HIR (8,22). The SNR of the lung parenchyma obtained by the DLIR variants was lower than that obtained by the ASIR algorithms. The superior CNR of the MPP lesion achieved by DLIR-H offsets this difference in the SNR, as it prioritizes the contrast between MPP lesions and normal lung tissue (23), which is considerably more relevant for clinical diagnosis than the global parenchymal SNR (8,24,25). These findings align with the clinical requirements for pediatric MPP imaging, where detecting subtle lesions (e.g., patchy GGOs) takes precedence over general parenchymal noise metrics (26).

Pediatric patients are radiation-sensitive (27). The radiation dose parameters (CTDI_vol: 2.97 mGy; ED: 1.17 mSv) used in this study are consistently below the reference levels recommended in three Chinese pediatric CT guidelines, confirming the compliance of our low-dose protocol. By validating the ability of DLIR-H to maintain the diagnostic quality at these low doses, we address the critical trade-off between dose reduction and image quality, which has restricted the clinical application of aggressive low-dose CT in children.

The high observer consistency of DLIR-H ensures reproducible diagnoses, which is key for clinical adoption. ASIR-80% images exhibited mild “waxy” artifacts that slightly obscured fine structures (28), which is consistent with previous reports of HIR limitations (9,11). In pediatric MPP, where lesions are often subtle and anatomical structures are smaller, the reduced number of artifacts observed in DLIR-H images directly improved the diagnostic confidence, as reflected in the highest subjective scores obtained for DLIR-H and the preference of radiologists for DLIR-H for visualizing GGO.

This study offers three innovations: first, it focuses on pediatric MPP, addressing its unique imaging challenges, including subtle parenchymal changes and the need for minimal radiation exposure. Second, objective metrics, subjective scoring, and observer agreement analysis are integrated for multidimensional comprehensive evaluation to validate the superiority of DLIR-H. Third, demonstrates DLIR-H’s utility at a median ED of 1.17 mSv is demonstrated, supporting further dose reduction beyond the required for ASIR.

Despite the strengths of this study, several limitations warrant discussion. First, as a retrospective study, we could not standardize breath-holding training for all patients in clinical practice, which may have affected the lung volume during scanning. Future prospective studies should incorporate standardized scanning protocols and objective respiratory motion assessments to minimize this variability. Second, we did not explore the maximum dose reduction achievable with DLIR-H while maintaining diagnostic quality. Dose-escalation studies could define the optimal radiation dose range for DLIR-H in pediatric patients with MPP, further enhancing its clinical utility. Third, the scanning equipment, and the ASIR and DLIR algorithms used in this study were proprietary to GE Healthcare, limiting the generalizability of their results when they are acquired from other manufacturers. Multicentre studies involving imaging with different CT scanners and reconstruction algorithms are needed to validate the performance of DLIR-H across platforms. Fourth, while we assessed the image quality and the confidence of radiologists, we did not evaluate whether compared with ASIR, DLIR-H improves the diagnostic sensitivity or specificity of MPP. Future research should include a diagnostic performance analysis with clinical follow-up as the reference standard.

Conclusions

DLIR-H outperforms ASIR variants in low-dose pediatric chest CT for MPP by minimizing background noise, maximizing the lesion CNR, preserving the natural tissue texture, and providing high-quality images with a clinically acceptable low radiation dose. It enhances the visualization of subtle MPP lesions and the diagnostic confidence of radiologists. By adhering to the STROBE guidelines and addressing methodological rigor, this study provides robust evidence for the clinical translation of DLIR-H in pediatric thoracic imaging. Future prospective multicentre studies should validate its diagnostic accuracy and generalizability across different CT platforms, further establishing DLIR-H as a standard reconstruction strategy for low-dose CT in pediatric MPP.

Acknowledgments

The authors thank Luotong Wang and Xu Yan, Senior engineers, GE for their supports for chest CT image post-processing and reconstruction.

Footnote

Reporting Checklist: The authors have completed the STROBE reporting checklist. Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-1-923/rc

Data Sharing Statement: Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-1-923/dss

Peer Review File: Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-1-923/prf

Funding: This work was supported by Tianjin Education Commission scientific research project (No. 2023KJ165) and Innovation Incubation Project of First Teaching Hospital of Tianjin University of Traditional Chinese Medicine (No. YuanZZ2024004).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-1-923/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. This study was approved by the Medical Ethics Committee of the First Teaching Hospital of Tianjin University of Traditional Chinese Medicine (No. TYLL2024[Z] 025), and the informed consent was waived because of the retrospective design.

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. Li Y, Wang X, Blau DM, et al. Global, regional, and national disease burden estimates of acute lower respiratory infections due to respiratory syncytial virus in children younger than 5 years in 2019: a systematic analysis. Lancet 2022;399:2047-64. [Crossref] [PubMed]
2. Wang Y, Yu X, Liu F, et al. Respiratory microbiota imbalance in children with Mycoplasma pneumoniae pneumonia. Emerg Microbes Infect 2023;12:2202272. [Crossref] [PubMed]
3. Global, regional, and national burden of 12 mental disorders in 204 countries and territories, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Psychiatry. 2022;9:137-50.
4. Li ZJ, Zhang HY, Ren LL, et al. Etiological and epidemiological features of acute respiratory infections in China. Nat Commun 2021;12:5026. [Crossref] [PubMed]
5. Debray MP, Carette MF, Loubet P, et al. CT features of community-acquired pneumonia at the emergency department. Respir Med Res 2022;81:100892. [Crossref] [PubMed]
6. Kuo BJ, Lai YK, Tan MLM, et al. Utility of Screening Chest Radiographs in Patients with Asymptomatic or Minimally Symptomatic COVID-19 in Singapore. Radiology 2021;298:E131-40. [Crossref] [PubMed]
7. Wang H, Li LL, Shang J, et al. Application of deep learning image reconstruction in low-dose chest CT scan. Br J Radiol 2022;95:20210380. [Crossref] [PubMed]
8. Zhang K, Shi X, Xie SS, et al. Deep learning image reconstruction in pediatric abdominal and chest computed tomography: a comparison of image quality and radiation dose. Quant Imaging Med Surg 2022;12:3238-50. [Crossref] [PubMed]
9. Carey S, Kandel S, Farrell C, et al. Comparison of conventional chest x ray with a novel projection technique for ultra-low dose CT. Med Phys 2021;48:2809-15. [Crossref] [PubMed]
10. Liu P, Wang M, Wang Y, et al. Impact of Deep Learning-based Optimization Algorithm on Image Quality of Low-dose Coronary CT Angiography with Noise Reduction: A Prospective Study. Acad Radiol 2020;27:1241-8. [Crossref] [PubMed]
11. Sun J, Li H, Li J, et al. Improving the image quality of pediatric chest CT angiography with low radiation dose and contrast volume using deep learning image reconstruction. Quant Imaging Med Surg 2021;11:3051-8. [Crossref] [PubMed]
12. Sato M, Ichikawa Y, Domae K, et al. Deep learning image reconstruction for improving image quality of contrast-enhanced dual-energy CT in abdomen. Eur Radiol 2022;32:5499-507. [Crossref] [PubMed]
13. 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]
14. Trattner S, Halliburton S, Thompson CM, et al. Cardiac-Specific Conversion Factors to Estimate Radiation Effective Dose From Dose-Length Product in Computed Tomography. JACC Cardiovasc Imaging 2018;11:64-74. [Crossref] [PubMed]
15. Radiology group of Chinese Pediatrics Society of Chinese Medical Association. Pediatrics group of Chinese Society of Radiology of Chinese Medical Association. Expert consensus on radiation dose of children CT examination. Chinese Journal of Radiology 2024;58:158-64.
16. Sun JH, Yu T, Li HY, et al. Investigation of radiation dose and diagnostic reference levels in CT scanning for children. Chin J Radiol 2022;56:1135-40.
17. Tuberculosis Group of the Professional Committee of Infectious Disease Imaging Technology RTBotCMA, Imaging Professional Committee of the Tuberculosis Society of the Chinese Medical Association. Expert Consensus on Standardization of Chest CT Scanning. China Med Devices 2020;35:185-9.
18. Sun J, Li H, Gao J, et al. Performance evaluation of a deep learning image reconstruction (DLIR) algorithm in "double low" chest CTA in children: a feasibility study. Radiol Med 2021;126:1181-8. [Crossref] [PubMed]
19. Sun J, Li H, Li H, et al. Application of deep learning image reconstruction algorithm to improve image quality in CT angiography of children with Takayasu arteritis. J Xray Sci Technol 2022;30:177-84. [Crossref] [PubMed]
20. Jensen CT, Liu X, Tamm EP, et al. Image Quality Assessment of Abdominal CT by Use of New Deep Learning Image Reconstruction: Initial Experience. AJR Am J Roentgenol 2020;215:50-7. [Crossref] [PubMed]
21. Tatsugami F, Higaki T, Nakamura Y, et al. Deep learning-based image restoration algorithm for coronary CT angiography. Eur Radiol 2019;29:5322-9. [Crossref] [PubMed]
22. Haggerty JE, Smith EA, Kunisaki SM, et al. CT imaging of congenital lung lesions: effect of iterative reconstruction on diagnostic performance and radiation dose. Pediatr Radiol 2015;45:989-97. [Crossref] [PubMed]
23. McCollough CH, Yu L, Kofler JM, et al. Degradation of CT Low-Contrast Spatial Resolution Due to the Use of Iterative Reconstruction and Reduced Dose Levels. Radiology 2015;276:499-506. [Crossref] [PubMed]
24. Tang H, Liu Z, Hu Z, et al. Clinical value of a new generation adaptive statistical iterative reconstruction (ASIR-V) in the diagnosis of pulmonary nodule in low-dose chest CT. Br J Radiol 2019;92:20180909. [Crossref] [PubMed]
25. Ng CKC. Artificial Intelligence for Radiation Dose Optimization in Pediatric Radiology: A Systematic Review. Children (Basel) 2022;9:1044. [Crossref] [PubMed]
26. Jensen CT, Wagner-Bartak NA, Vu LN, et al. Detection of Colorectal Hepatic Metastases Is Superior at Standard Radiation Dose CT versus Reduced Dose CT. Radiology 2019;290:400-9. [Crossref] [PubMed]
27. 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]
28. Brady SL, Trout AT, Somasundaram E, et al. Improving Image Quality and Reducing Radiation Dose for Pediatric CT by Using Deep Learning Reconstruction. Radiology 2021;298:180-8. [Crossref] [PubMed]