Prospective Comparison of Ultra-Low-Dose Versus Standard-Dose Computed Tomography for Esophageal Foreign Body Detection

Summary

This randomized trial demonstrates that ultra-low-dose computed tomography (CT) with deep learning reconstruction achieves diagnostic accuracy non-inferior to that of standard-dose CT for esophageal foreign body detection, with a 63% reduction in radiation dose.

Abstract

The objective of this study was to verify the non-inferiority of ultra-low-dose computed tomography (ULD-CT) versus standard-dose CT (SD-CT) for esophageal foreign body (EFB) detection and to quantify dose-quality trade-offs. This prospective, randomized, blinded-reader study enrolled 180 patients (120 adults, 60 children) presenting with suspected EFB ingestion. Patients were randomized to ULD-CT (100 kV/50 mA, adaptive statistical iterative reconstruction-V 80% + deep learning image reconstruction) or SD-CT (120 kV/200 mA, filtered back projection + adaptive statistical iterative reconstruction-V 30%) groups. All patients underwent endoscopic or surgical reference-standard evaluation within 12 h of CT imaging, with 30-day follow-up for negative findings. The primary endpoint was the area under the curve (AUC); secondary endpoints included image quality metrics, radiation dose, and incidental findings.

Both protocols achieved excellent diagnostic performance, with sensitivity/specificity of 97.8%/100% for ULD-CT and 98.9%/100% for SD-CT. The AUC was 0.985 for ULD-CT versus 0.991 for SD-CT (difference < non-inferiority margin). The effective dose was reduced by 63% with ULD-CT (P < 0.001). The signal-to-noise ratio decreased by 37% (P < 0.001), yet diagnostic acceptability (score ≥3) was maintained at 94.4% versus 97.8% (P = 0.070). Incidental findings were identified in 54 patients (30.0%), including 19 potentially actionable findings (10.6%) and 6 high-significance findings requiring urgent evaluation.

Ultra-low-dose CT demonstrates non-inferior diagnostic accuracy compared with SD-CT for EFB detection while reducing radiation exposure by over 60%. These findings show potential for ULD-CT as a first-line modality for suspected EFBs pending multicenter validation.

Introduction

Foreign body ingestion remains a common emergency presentation worldwide, with esophageal foreign bodies (EFBs) accounting for a clinically important subset that often requires urgent imaging or endoscopic intervention^1,2,3. Recent radiology reviews emphasize that CT is particularly valuable when radiographs are negative, the object is radiolucent, or complications such as perforation are suspected^1,2. Special patient populations, including individuals with intellectual disabilities or impaired communication, may present late or with atypical symptoms and may ingest non-food, sharp-edged, or multiple objects, creating additional diagnostic and management challenges⁴. Recent hospital-based studies from China further show that sharp foreign bodies and food impactions remain common causes of EFB presentation in adults and older patients, although nationwide incidence data are still lacking^5,6.

Despite the clinical importance of accurate EFB detection, major knowledge gaps remain regarding optimal imaging strategies. Current protocols typically employ standard radiation doses developed for general thoracic imaging rather than indication-specific parameters optimized for foreign body conspicuity, while dose-reduction guidance emphasizes tailoring exposure to the diagnostic task⁷. Furthermore, although iterative reconstruction and deep learning reconstruction have enabled marked radiation reductions in other CT applications, their impact on EFB detection has not been systematically evaluated in prospective randomized trials⁸.

Traditional imaging evaluation of suspected EFBs has relied on plain radiography followed by selective contrast esophagography or endoscopy for radiolucent objects^1,3,9. However, computed tomography (CT) has increasingly emerged as the preferred modality due to its superior sensitivity for both radiopaque and radiolucent foreign bodies. Studies have demonstrated that CT sensitivity approaches 100% compared with 70%–80% for plain radiographs, particularly for fishbones and other low-density materials^10,11. The cross-sectional nature of CT also provides crucial information about complications, including perforation, abscess formation, and vascular proximity, that may alter patient management^12,13.

The widespread adoption of CT has raised concerns about cumulative radiation exposure, particularly in pediatric populations, in which lifetime cancer risk from medical imaging is highest^14,15. Traditional CT protocols deliver effective doses of 8–12 mSv for chest examinations, prompting the development of dose reduction strategies aligned with the "as low as reasonably achievable" principle and modern CT dose-management recommendations^7,16.

Recent technological advances have enabled dramatic dose reductions while maintaining diagnostic quality. Iterative reconstruction algorithms such as adaptive statistical iterative reconstruction reduce image noise compared with filtered back projection, allowing lower radiation doses¹⁷. The newest generation of deep learning image reconstruction (DLIR) algorithms leverage convolutional neural networks trained on high-quality datasets to distinguish signal from noise, achieving superior noise reduction to conventional iterative techniques^8,18. Studies have demonstrated that DLIR can maintain diagnostic image quality at radiation doses substantially lower than those used in standard protocols^19,20.

Despite the promise of ultra-low-dose CT (ULD-CT) protocols, their application specifically for EFB detection remains unexplored in prospective randomized trials. Previous studies of low-dose chest CT have focused on lung nodule detection or general thoracic pathology rather than the unique diagnostic requirements of foreign body evaluation^21,22. Additionally, the impact of dose reduction on the detection of clinically relevant incidental findings—a potential added value of CT imaging—has not been systematically evaluated in this context²³.

This study aimed to address these knowledge gaps by conducting a prospective comparison of ULD-CT and standard-dose CT (SD-CT) for EFB detection. We hypothesized that (1) ULD-CT would demonstrate non-inferiority to SD-CT for diagnostic accuracy (non-inferiority margin δ = 5%); (2) the radiation dose would be substantially reduced while maintaining acceptable image quality; and (3) the broader field-of-view advantage of CT would enable detection of clinically relevant incidental findings that could impact patient management beyond the presenting complaint.

Protocol

This prospective, randomized, blinded-reader study was conducted at a tertiary care center (The First Hospital of Hebei Medical University) between January 2024 and June 2025. This study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of The First Hospital of Hebei Medical University (Approval Number: [2025]YS-062), and all patients or their legal guardians provided written informed consent prior to enrollment. For minors, informed consent was obtained according to age and capacity: for participants under 8 years of age, written informed consent was obtained from their legal guardians, with additional assent sought from the child when capable of understanding; for participants aged 8–17 years, written informed consent was obtained from both the minor and their legal guardian, with information disclosed at an age-appropriate level; participants aged 16–17 years who demonstrated financial independence through their own labor income were considered to have full civil capacity and could provide independent consent. The study has been registered with the UK Clinical Study Registry, registration number ISRCTN13588740.

Study design
Patient randomization was performed using a computer-generated sequence with permuted blocks of varying sizes (4, 6, and 8) stratified by age group (pediatric < 18 years vs. adult ≥ 18 years). Allocation concealment was maintained through sealed opaque envelopes opened immediately prior to CT scanning. All image interpreters, endoscopists, and surgeons were blinded to the CT protocol assignment.

Study population
Inclusion criteria
Eligible participants were patients aged 3–80 years presenting to the emergency department and referred for CT evaluation based on institutional protocols (primarily for suspected radiolucent foreign bodies not visible on plain radiography, clinical suspicion of complications, or need for precise localization prior to intervention), with (1) a history of foreign body ingestion within 6 hours of presentation; (2) symptoms suggestive of esophageal impaction, including dysphagia, odynophagia, chest pain, or hypersalivation; and (3) planned endoscopic or surgical evaluation within 12 h of CT imaging. The 6 h window was selected to ensure patients represented acute presentations while allowing sufficient time for imaging and endoscopic evaluation, though this may limit generalizability to delayed presentations common with certain foreign body types, such as fishbones.

Exclusion criteria
Patients were excluded if they had (1) severe cardiorespiratory instability requiring immediate intervention; (2) pregnancy or a positive pregnancy test; (3) a known contrast allergy (for enhanced scans when clinically indicated); (4) prior esophageal surgery or known esophageal stricture; (5) metallic implants causing substantial artifacts affecting >30% of the esophageal evaluation area; and (6) body mass index (BMI) > 40 kg/m² (due to potential image quality degradation at ultra-low doses).

Computed tomography scanning protocols
All examinations were performed on 256-slice multi-detector CT scanners with deep-learning reconstruction capabilities. Patients were positioned supine with their arms elevated above their heads when possible. No oral contrast was administered to avoid obscuring foreign bodies or delaying endoscopy.

The SD-CT protocol utilized parameters consistent with the routine chest CT protocol at our institution: a tube voltage of 120 kV, a reference tube current of 200 mA with automatic tube current modulation (ATCM) enabled, a rotation time of 0.5 s, a pitch of 0.992, and collimation of 0.625 mm. Images were reconstructed using filtered back projection with 30% adaptive statistical iterative reconstruction blending, representing the current clinical standard at the participating site. The rationale for 120 kV was based on standard thoracic imaging protocols optimized for general diagnostic purposes.

The ULD-CT protocol employed aggressive dose reduction strategies: a tube voltage of 100 kV, a reference tube current of 50 mA with ATCM (range 10–80 mA), and an identical rotation time and pitch to that of the SD protocol. The 100 kV/50 mA parameters were selected based on preliminary phantom studies demonstrating maintained foreign body conspicuity at these settings when combined with advanced reconstruction. Raw data were reconstructed using 80% adaptive statistical iterative reconstruction blending, followed by DLIR at medium strength. This dual-reconstruction approach maximized noise reduction while preserving anatomical detail and avoiding the plastic appearance sometimes associated with aggressive iterative reconstruction alone.

For both protocols, images were reconstructed at a 0.625 mm slice thickness with a 0.5 mm overlap to enable multiplanar reformations. The scan range extended from the lower neck (C3 level) through the gastroesophageal junction, with careful positioning to minimize breast tissue inclusion in female patients. Dose reduction features, including organ-based tube current modulation and adaptive collimation, were enabled for all scans.

Data collection and variables
Patient characteristics
Demographic and clinical data collected included age, sex, BMI, presenting symptoms, time from ingestion to imaging, type of foreign body reported by history, relevant comorbidities (diabetes mellitus, chronic obstructive pulmonary disease, prior thoracic malignancy), and prior CT examinations within 6 months.

Foreign body characteristics
For confirmed EFB cases, the following were documented: (1) material composition (bone, metal, food bolus, plastic, other); (2) maximum dimension measured on CT; (3) attenuation in Hounsfield units (HU) measured using a standardized 5 mm² region of interest; (4) anatomical location using established landmarks (cervical C3–C7, upper thoracic T1–T4, mid-thoracic T5–T8, lower thoracic T9–T12); and (5) the presence of complications, including perforation, pneumomediastinum, or abscess formation.

Image quality assessment
Objective image quality metrics were measured by a medical physicist blinded to the protocol assignment. The signal-to-noise ratio (SNR) was calculated as the mean attenuation of the descending aorta divided by the standard deviation of subcutaneous fat. The contrast-to-noise ratio was calculated as the difference in attenuation between aortic blood and paraspinal muscle divided by image noise. To better align with the diagnostic task, additional measurements were performed in the paraesophageal fat. Measurements were performed on axial images at three standardized levels (aortic arch, carina, and mid-esophagus) with circular regions of interest (150 mm²) placed consistently using anatomical landmarks.

Subjective image quality was independently assessed by two thoracic radiologists with 8 years and 12 years of experience. Images were reviewed on diagnostic workstations using standardized soft-tissue and lung window settings (window width: 350 HU, window level: 40 HU for soft tissue; window width: 1,500 HU, window level: -600 HU for lung evaluation), consistent with routine thoracic CT interpretation and artifact-recognition principles²⁴. Readers scored the following parameters on a 5-point Likert scale: (1) edge definition of the mediastinal structures; (2) image noise; (3) diagnostic confidence for foreign body detection; and (4) overall diagnostic quality. A score ≥3 was considered diagnostically acceptable. Discrepancies between readers were resolved through consensus review, with the consensus score used for analysis. Inter-reader agreement was assessed using weighted kappa statistics.

Radiation dose metrics
The scanner-reported volume CT dose index (CTDIvol) and dose-length product were recorded for each examination. The effective dose (ED) was calculated using age- and sex-specific conversion factors (k = 0.014 mSv·mGy⁻¹·cm⁻¹ for adults; age-adjusted factors for pediatric patients) based on International Commission on Radiological Protection Publication 103 recommendations²⁵. Size-specific dose estimates were calculated using patient anteroposterior and lateral dimensions measured at the mid-chest level.

Reference standard
The reference standard for foreign body presence and location was established through endoscopic visualization or surgical findings performed within 12 h of CT imaging for all randomized patients, irrespective of CT protocol assignment or CT result. Endoscopy reports were reviewed by two gastroenterologists to confirm the foreign body characteristics and anatomical location using standardized landmarks. When endoscopy or surgery did not identify a retained foreign body, clinical follow-up at 30 days through a chart review and telephone contact confirmed the absence of missed foreign bodies, with specific inquiry about return visits, delayed complications, or the need for repeat imaging or endoscopy. A uniform reference-standard application was used to minimize differential verification bias.

Incidental findings
All CT examinations were systematically reviewed for incidental findings unrelated to the indication for imaging by the same two radiologists who performed the quality assessment. The findings were categorized by anatomical location (pulmonary, mediastinal, cardiovascular, upper abdominal, osseous, other) and clinical significance. Clinical significance was classified as follows: (1) low—findings requiring no follow-up (e.g., simple hepatic cysts, degenerative spine changes); (2) moderate—findings potentially requiring follow-up imaging (e.g., thyroid nodules > 1 cm, indeterminate adrenal nodules); (3) high—findings requiring urgent evaluation or intervention (e.g., suspicious pulmonary nodules, aortic aneurysm > 5 cm, suspicious breast masses). Age and sex distributions of the incidental findings were recorded, and downstream management pathways were documented for all actionable findings.

Statistical analysis
Sample size calculation
Sample size was calculated based on the primary endpoint of diagnostic accuracy (the area under the curve; AUC). Assuming a standard-dose AUC of 0.97 based on the literature and an expected ULD-CT AUC of 0.95, a 5-percentage-point non-inferiority margin was selected a priori because an AUC loss greater than 0.05 would be clinically meaningful enough to alter imaging triage, whereas a smaller reduction was considered acceptable when balanced against substantial radiation reduction and mandatory endoscopic or surgical confirmation. Using a one-sided alpha of 0.025, 80% power, independent patient groups, and ROC-based sample size assumptions for diagnostic accuracy studies²⁶, 166 patients were required. Accounting for an 8% dropout rate or technical failure, the target enrollment was 180 patients.

Primary analysis
The primary analysis compared the area under the receiver operating characteristic (ROC) curve between ULD-CT and SD-CT for foreign body detection. Non-inferiority was declared if the lower bound of the 95% confidence interval (CI) for the AUC difference (ULD-CT minus SD-CT) exceeded -0.05, and ROC curves were constructed using radiologist confidence scores (1–5 scale) as the diagnostic variable. Because this was a parallel-arm design with independent patient groups, the DeLong method for independent ROC-curve comparison was used for statistical testing²⁷. Reader scores were averaged when both readers provided assessments, and this average was used as the diagnostic variable for ROC analysis. Sensitivity and specificity denominators represent reference-standard positive and reference-standard negative diagnostic decision units, respectively, rather than the total number of randomized participants in each protocol arm.

Secondary analyses
Given the parallel-arm randomized design, all comparative analyses used independent-sample methods. Continuous variables (dose metrics, image quality scores) were compared between protocols using independent-sample t-tests or Mann-Whitney U tests based on distribution normality assessed by Shapiro-Wilk testing. Categorical variables were compared using chi-square tests or Fisher's exact tests for unpaired proportions. Sensitivity, specificity, and accuracy were compared using Wald CIs and chi-square tests for independent samples. Inter-reader agreement was assessed using weighted kappa statistics with quadratic weights within each protocol arm separately.

Subgroup analyses examined diagnostic performance stratified by (1) age group (pediatric vs. adult); (2) foreign body density (high > 100 vs. low ≤ 100 HU); (3) BMI categories (<25, 25–30, >30 kg/m²); and (4) anatomical location (cervical vs. thoracic esophagus). Interaction terms were formally tested using logistic regression models.

Multivariable logistic regression identified predictors of missed or indeterminate foreign bodies, with candidate variables including foreign body size, density, anatomical location, patient BMI, image noise (SNR), and reconstruction algorithm. Interaction terms for protocol × density and protocol × BMI were included based on a priori hypotheses. Model selection used backward elimination with a retention threshold of P < 0.10.

For incidental findings, because patients were randomized to different protocols and did not undergo both scans, sensitivity and agreement calculations across protocols were not appropriate and were removed from the analysis. Instead, we compared the prevalence and distribution of incidental findings between protocols using chi-square tests.

Contrast-enhanced examinations (performed in 23 patients in the SD-CT arm and 21 patients in the ULD-CT arm based on a clinical indication for vascular or mediastinal evaluation) were analyzed separately to assess the influence on diagnostic confidence and incidental finding detection.

All analyses followed intention-to-treat principles. Missing data (<2% overall) were handled using multiple imputation with 10 imputed datasets. Statistical analyses were performed using R version 4.3.2 and MedCalc version 20.0. Two-sided P-values < 0.05 were considered statistically significant except in the primary non-inferiority analysis.

Results

Patient demographics and scan parameters
Between January 2024 and June 2025, 198 patients were assessed for eligibility (Supplementary Figure 1). Eighteen patients were excluded (8 due to immediate surgical indications, 6 with BMI > 40 kg/m², and 4 who declined consent), leaving 180 patients randomized equally to the SD-CT and ULD-CT protocols. All patients completed their assigned CT protocol and reference standard evaluation. Table 1 summarizes the baseline demographics and scan parameters. The groups were well balanced, with no significant differences in age, sex, BMI, or foreign body types. The ULD protocol achieved a 63% reduction in ED (1.88 mSv vs. 5.09 mSv, P < 0.001). Contrast-enhanced CT was performed in 44 patients (24.4%) based on clinical indication (suspected perforation or vascular complications): 23 (25.6%) in the SD-CT arm and 21 (23.3%) in the ULD-CT arm (P = 0.729).

Image quality metrics
Table 2 presents a comprehensive image quality assessment. Although objective noise metrics were significantly higher with ULD-CT, diagnostic acceptability (score ≥ 3) was maintained in 94.4% of cases, which was not significantly different from SD-CT (97.8%, P = 0.070). Inter-reader agreement remained excellent for both protocols (weighted kappa = 0.81 for ULD-CT, 0.84 for SD-CT, P = 0.528 for the comparison between protocols). Figure 1 presents representative CT images from two confirmed cases—a cervical date-pit impaction and a lower-esophageal bezoar—demonstrating that clinically actionable foreign bodies remained conspicuous on routine multiplanar reconstruction images.

Diagnostic performance
The primary endpoint analysis demonstrated the non-inferiority of ULD-CT to SD-CT (Table 3). The AUC was 0.985 (95% CI: 0.969–0.997) for ULD-CT versus 0.991 (95% CI: 0.978–0.999) for SD-CT, with a difference of −0.006 (95% CI: −0.018 to 0.006), satisfying the prespecified non-inferiority margin of -0.05. Both protocols achieved excellent sensitivity (97.8% vs. 98.9%) and perfect specificity (100%).

Subgroup analyses revealed maintained performance across patient populations (Table 3). Both protocols achieved perfect diagnostic accuracy in pediatric patients, with sensitivity and specificity of 100% in both arms and 95% CIs of 88.8%–100% for sensitivity and 87.5%–100% for specificity. For low-density foreign bodies (≤100 HU), including food boluses and plastic objects, ULD-CT showed slightly reduced sensitivity compared with SD-CT (93.8% [30/32] vs. 96.9% [31/32]), with overlapping 95% CIs (79.2%–98.2% vs. 83.8%–99.4%).

The time from ingestion to imaging did not significantly affect diagnostic performance in either protocol (stratified analysis at <3 h vs. 3–6 h showed no significant differences, P = 0.623).

Figure 2 displays the ROC curves comparing diagnostic performance. Both curves demonstrate excellent discriminatory ability with minimal separation, supporting the non-inferiority conclusion. The AUC values of 0.991 for SD-CT and 0.985 for ULD-CT confirm that both protocols achieve near-perfect classification performance.

Radiation dose analysis
The ULD protocol achieved a consistent dose reduction of approximately 63% across all patient subgroups (Table 4). The absolute dose reduction was most pronounced in patients with obesity (10.7 mGy reduction in CTDIvol) while maintaining a similar percentage reduction. Organ dose estimates showed proportional reductions, with the thyroid dose decreasing from 8.2 mGy to 3.0 mGy—particularly relevant given the radiosensitivity of this organ. Based on the biological effects of ionizing radiation (BEIR) VII models, the lifetime attributable cancer risk was reduced from 74.2 to 27.4 per 100,000 exposed 10-year-old children, representing the prevention of approximately 47 theoretical cancers per 100,000 pediatric CT examinations. These estimates should be interpreted cautiously, given the known limitations and uncertainties of low-dose radiation risk models, particularly for individual risk prediction.

Figure 3 illustrates the relationship between radiation dose and image quality across individual examinations. A clear separation between the two protocols is evident, with SD-CT clustered around 5–8 mSv, achieving SNR values of 15–30, and ULD-CT centered around 1–3 mSv with SNR values of 5–20. The horizontal dashed line at SNR = 8 indicates the proposed diagnostic quality threshold based on the ROC analysis. Despite lower SNR values, the majority of ULD-CT examinations remained above this threshold, demonstrating that diagnostic quality can be maintained at substantially reduced radiation doses.

Incidental findings
Incidental findings were detected in 54 patients (30.0%), with no significant difference in prevalence between protocols (31.1% for SD-CT vs. 28.9% for ULD-CT, P = 0.745) (Table 5). Most findings (64.8%) were of low clinical significance, requiring no follow-up. However, 19 findings (10.6% of patients) were deemed potentially actionable, including 6 requiring urgent evaluation. Among these, two malignancies were diagnosed and treated (one thyroid cancer in the SD-CT arm, one breast cancer in the ULD-CT arm) that would have otherwise remained undetected.

All high-significance findings requiring urgent evaluation were detected by both protocols when present in the respective patient groups. The distribution of incidental findings by clinical significance showed no statistically significant differences between protocols (P = 0.851 for low significance, P = 0.773 for moderate significance, P = 1.000 for high significance).

Predictors of diagnostic challenges
Multivariable analysis identified key predictors of diagnostic challenges (Table 6). Small foreign body size (<10 mm) and low density (≤100 HU) were the strongest predictors, with odds ratios (ORs) of 3.42 (95% CI: 1.28–9.14, P = 0.014) and 2.87 (95% CI: 1.09–7.56, P = 0.033), respectively. Image quality, quantified by SNR < 10, was also strongly associated with diagnostic difficulty (OR: 4.68, 95% CI: 1.73–12.66, P = 0.002). The ULD protocol itself was not an independent predictor of missed cases (OR: 1.42, 95% CI: 0.48–4.21, P = 0.527), although there was a trend toward interaction with low-density foreign bodies (OR: 2.94, 95% CI: 0.87–9.93, P = 0.083), suggesting that these cases may benefit from SD imaging. As reported in Table 6, the model achieved good discrimination (C-statistic = 0.84, 95% CI: 0.76–0.92).

Figure 4 presents a forest plot summarizing diagnostic accuracy across key subgroups. Both protocols maintain excellent accuracy (>94%) across all subgroups, with CIs demonstrating the precision of the estimates. The minimal difference between the protocols is most apparent for low-density foreign bodies, where ULD-CT shows slightly wider CIs while still achieving clinically acceptable accuracy. The vertical dashed line at 95% represents the non-inferiority threshold, with all subgroup estimates exceeding this benchmark.

DATA AVAILABILITY
The de-identified raw/source data supporting the reported tables and figures are provided in Supplementary File 1.

Figure 1: Representative author-supplied CT images of confirmed esophageal foreign bodies. (A, B) A 12-year-old boy with a date pit impacted at the esophageal inlet. Axial and coronal CT images show a spindle-shaped hyperdense foreign body traversing the upper esophageal lumen with both sharp ends closely apposed to the adjacent soft tissues. Endoscopy confirmed a date pit with mucosal injury at approximately 15 cm from the incisors. (C–E) A 62-year-old man with a lower-esophageal bezoar. Axial, coronal, and sagittal CT images show a round mixed-density intraluminal lesion in the distal esophagus measuring approximately 2.85 cm, consistent with an impacted bezoar causing luminal obstruction. Endoscopy confirmed a hard bezoar lodged approximately 30 cm from the incisors. Please click here to view a larger version of this figure.

Figure 2: Receiver operating characteristic (ROC) curves for foreign body detection. Both curves demonstrate excellent discriminatory ability with area under the curve (AUC) values of 0.991 (95% CI: 0.978–0.999) for standard-dose CT (blue) and 0.985 (95% CI: 0.969–0.997) for ultra-low-dose CT (red). The minimal separation between curves supports the non-inferiority conclusion. The dashed reference line indicates the non-inferiority margin of 0.05. Please click here to view a larger version of this figure.

Figure 3: Scatter plot showing the relationship between effective dose (ED) and signal-to-noise ratio (SNR) for individual examinations. Standard-dose CT (blue circles, n = 90) clusters around 5-8 mSv with SNR 15–30, while ultra-low-dose CT (red triangles, n = 90) centers around 1–3 mSv with SNR 5–20. Linear regression analysis shows: SD-CT: SNR = 3.42 × ED + 2.85 (R² = 0.81, 95% CI: 3.15–3.69); ULD-CT: SNR = 2.73 × ED + 4.12 (R² = 0.68, 95% CI: 2.38–3.08). The horizontal dashed line at SNR = 8 indicates the proposed diagnostic quality threshold. Regression lines show the dose-SNR relationship for each protocol. Despite lower SNR values, most ULD-CT examinations exceed the diagnostic threshold, demonstrating maintained quality at reduced dose. Please click here to view a larger version of this figure.

Figure 4: Forest plot showing diagnostic accuracy (with 95% confidence intervals) across patient and foreign body subgroups. Standard-dose CT (blue circles) and ultra-low-dose CT (red triangles) maintain excellent accuracy (>94%) across all subgroups. Sample sizes are indicated in parentheses. The vertical dashed line at 95% represents the non-inferiority threshold. Interaction P-values: protocol × age P = 0.623; protocol × density P = 0.083; protocol × BMI P = 0.407. Please click here to view a larger version of this figure.

Characteristic	Standard-Dose CT (n = 90)	Ultra-Low-Dose CT (n = 90)	P-value
Demographics
Age, years (median, IQR)	45 (28–62)	43 (26–59)	0.716
- Adults, n (%)	60 (66.7)	60 (66.7)	1.000
- Children, n (%)	30 (33.3)	30 (33.3)	1.000
Male sex, n (%)	48 (53.3)	51 (56.7)	0.653
BMI, kg/m² (mean ± SD)	26.4 ± 5.2	25.9 ± 5.6	0.536
Foreign Body Type, n (%)
Bone (fish/chicken)	38 (42.2)	35 (38.9)	0.649
Food bolus	24 (26.7)	27 (30.0)	0.620
Coin/metal	12 (13.3)	10 (11.1)	0.649
Dental prosthesis	8 (8.9)	9 (10.0)	0.799
Other/unknown	8 (8.9)	9 (10.0)	0.799
Scan Parameters
Tube voltage, kV	120	100	<0.001
Reference mAs	200	50	<0.001
Actual mAs (mean ± SD)	186.4 ± 42.3	48.2 ± 12.6	<0.001
Scan length, cm (mean ± SD)	28.4 ± 3.2	28.6 ± 3.4	0.685
Dose Metrics
CTDIvol, mGy (mean ± SD)	12.8 ± 3.4	4.7 ± 1.2	<0.001
DLP, mGy·cm (mean ± SD)	363.5 ± 98.2	134.4 ± 36.7	<0.001
Effective dose, mSv (mean ± SD)	5.09 ± 1.37	1.88 ± 0.51	<0.001
SSDE, mGy (mean ± SD)	14.2 ± 3.8	5.2 ± 1.4	<0.001

Table 1: Baseline demographics and scan parameters. Abbreviations: IQR, interquartile range; SD, standard deviation; BMI, body mass index; CTDIvol, volume CT dose index; DLP, dose-length product; SSDE, size-specific dose estimate.

Parameter	Standard-Dose CT	Ultra-Low-Dose CT	Difference (95% CI)	P-value
Objective Measures (mean ± SD)
SNR	18.4 ± 4.2	11.6 ± 3.8	-6.8 (-7.6 to -6.0)	<0.001
CNR	14.2 ± 3.6	8.9 ± 3.2	-5.3 (-6.0 to -4.6)	<0.001
Image noise, HU	12.3 ± 2.8	19.5 ± 4.6	7.2 (6.1 to 8.3)	<0.001
Subjective Scores (median, IQR)
Edge definition	4 (4-5)	4 (3-4)	-	0.018
Image noise	4 (4-5)	3 (3-4)	-	<0.001
Diagnostic confidence	5 (4-5)	4 (4-5)	-	0.076
Overall quality	4 (4-5)	4 (3-4)	-	0.026
Diagnostic Acceptability
Score ≥ 3, n (%)	88 (97.8)	85 (94.4)	-3.3% (-8.9 to 2.2)	0.070
Score ≥ 4, n (%)	82 (91.1)	68 (75.6)	-15.6% (-26.0 to -5.1)	0.003
Inter-reader Agreement
Weighted κ (95% CI)	0.84 (0.78-0.90)	0.81 (0.74-0.88)	-	0.528

Table 2: Image quality metrics. Abbreviations: SD, standard deviation; IQR, interquartile range; SNR, signal-to-noise ratio; CNR, contrast-to-noise ratio; HU, Hounsfield units; CI, confidence interval. Continuous variables compared using independent t-tests; subjective scores compared using Mann-Whitney U tests; categorical variables compared using chi-square tests.

Performance Metric	Standard-Dose CT	Ultra-Low-Dose CT	Difference (95% CI)
Overall Performance
Sensitivity, % (TP/Npos)	98.9 (89/90)	97.8 (88/90)	-1.1 (-5.2 to 3.0)
- 95% CI	94.0–99.8	92.1–99.4
Specificity, % (TN/Nneg)	100 (90/90)	100 (90/90)	0 (0 to 0)
- 95% CI	95.1–100	95.1–100
PPV, %	100	100	0 (0 to 0)
- 95% CI	96.0–100	95.9–100
NPV, %	98.8	97.6	-1.2 (-5.4 to 3.0)
- 95% CI	93.3–99.8	91.6–99.3
Accuracy, %	99.4	98.9	-0.6 (-2.9 to 1.8)
- 95% CI	96.9–99.9	96.0–99.7
AUC (95% CI)	0.991 (0.978–0.999)	0.985 (0.969–0.997)	-0.006 (-0.018 to 0.006)*
Subgroup Analysis - Adults (n = 120)
Sensitivity, % (TP/Npos)	98.3 (59/60)	96.7 (58/60)	-1.6 (-7.8 to 4.5)
- 95% CI	90.9–99.7	88.5–99.1
Specificity, % (TN/Nneg)	100 (60/60)	100 (60/60)	0 (0 to 0)
- 95% CI	94.0–100	94.0–100
AUC (95% CI)	0.992 (0.975–0.999)	0.983 (0.962–0.996)	-0.009 (-0.024 to 0.006)
Subgroup Analysis - Children (n = 60)
Sensitivity, % (TP/Npos)	100 (30/30)	100 (30/30)	0 (0 to 0)
- 95% CI	88.8–100	88.8–100
Specificity, % (TN/Nneg)	100 (30/30)	100 (30/30)	0 (0 to 0)
- 95% CI	87.5–100	87.5–100
AUC (95% CI)	1.000 (0.985–1.000)	1.000 (0.985–1.000)	0 (0 to 0)
By Foreign Body Density
High density (>100 HU)
- Sensitivity, % (TP/Npos)	100 (50/50)	100 (45/45)	0 (0 to 0)
- 95% CI	93.5–100	93.5–100
- AUC (95% CI)	1.000 (0.987–1.000)	1.000 (0.987–1.000)	0 (0 to 0)
Low density (≤100 HU)
- Sensitivity, % (TP/Npos)	96.9 (31/32)	93.8 (30/32)	-3.1 (-13.0 to 6.8)
- 95% CI	83.8–99.4	79.2–98.2
- AUC (95% CI)	0.984 (0.952–0.998)	0.969 (0.932–0.992)	-0.015 (-0.042 to 0.012)

Table 3: Diagnostic performance for foreign body detection. *Non-inferiority criterion met (lower bound of 95% CI > -0.05). Abbreviations: CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value; AUC, area under the curve; HU, Hounsfield units; TP, true positives; TN, true negatives; Npos, reference-standard positive diagnostic decision units; Nneg, reference-standard negative diagnostic decision units. For sensitivity, the denominator is Npos; for specificity, the denominator is Nneg; these denominators are diagnostic decision units and should not be summed as the randomized participant count. Confidence intervals were calculated using the binomial exact method for sensitivity/specificity; the DeLong method was used for independent AUC comparison.

Subgroup	Standard-Dose CT	Ultra-Low-Dose CT	Absolute Reduction	Percentage Reduction
Overall
CTDIvol, mGy (mean ± SD)	12.8 ± 3.4	4.7 ± 1.2	8.1	63.3%
SSDE, mGy (mean ± SD)	14.2 ± 3.8	5.2 ± 1.4	9.0	63.4%
Effective dose, mSv (mean ± SD)	5.09 ± 1.37	1.88 ± 0.51	3.21	63.1%
By BMI Category
BMI <25 kg/m²
- CTDIvol, mGy	10.2 ± 2.8	3.8 ± 0.9	6.4	62.7%
- SSDE, mGy	11.3 ± 3.1	4.2 ± 1.0	7.1	62.8%
- Effective dose, mSv	4.08 ± 1.12	1.52 ± 0.36	2.56	62.7%
BMI 25-30 kg/m²
- CTDIvol, mGy	13.1 ± 3.2	4.8 ± 1.1	8.3	63.4%
- SSDE, mGy	14.5 ± 3.5	5.3 ± 1.2	9.2	63.4%
- Effective dose, mSv	5.24 ± 1.28	1.92 ± 0.44	3.32	63.4%
BMI >30 kg/m²
- CTDIvol, mGy	17.6 ± 4.1	6.9 ± 1.6	10.7	60.8%
- SSDE, mGy	19.5 ± 4.5	7.6 ± 1.8	11.9	61.0%
- Effective dose, mSv	7.04 ± 1.64	2.76 ± 0.64	4.28	60.8%
By Age Group
Adults (≥18 years)
- CTDIvol, mGy	13.5 ± 3.6	5.0 ± 1.3	8.5	63.0%
- SSDE, mGy	15.0 ± 4.0	5.5 ± 1.4	9.5	63.3%
- Effective dose, mSv	5.42 ± 1.41	2.01 ± 0.52	3.41	62.9%
Children (<18 years)
- CTDIvol, mGy	9.2 ± 2.5	3.4 ± 0.9	5.8	63.0%
- SSDE, mGy	10.2 ± 2.8	3.8 ± 1.0	6.4	62.7%
- Effective dose, mSv	3.67 ± 0.98	1.37 ± 0.36	2.30	62.7%
Organ Dose Estimates
Thyroid, mGy	8.2 ± 2.1	3.0 ± 0.8	5.2	63.4%
Breast, mGy	6.8 ± 1.8	2.5 ± 0.7	4.3	63.2%
Lung, mGy	9.4 ± 2.5	3.5 ± 0.9	5.9	62.8%
Lifetime Attributable Risk† (per 100,000)
10-year-old children	74.2	27.4	46.8	63.1%
30-year-old adults	42.8	15.8	27.0	63.1%
50-year-old adults	18.6	6.9	11.7	62.9%

Table 4: Radiation dose analysis by patient subgroups. Abbreviations: SD, standard deviation; BMI, body mass index; CTDIvol, volume CT dose index; SSDE, size-specific dose estimate. †Based on BEIR VII models for lifetime cancer incidence. Values represent theoretical estimates with inherent model uncertainties. Note: DLP to effective dose conversion factor k = 0.014 mSv·mGy⁻¹·cm⁻¹ for adults; age-adjusted factors for pediatric patients per ICRP 103.

Finding Category	Total n (%)	SD-CT n (%)	ULD-CT n (%)	P-value
Any Incidental Finding	54 (30.0)	28 (31.1)	26 (28.9)	0.745
By Anatomical Location
Pulmonary	23 (12.8)	12 (13.3)	11 (12.2)	0.823
- Nodules < 6 mm	14 (7.8)	7 (7.8)	7 (7.8)	1.000
- Nodules ≥ 6 mm	4 (2.2)	2 (2.2)	2 (2.2)	1.000
- Emphysema	3 (1.7)	2 (2.2)	1 (1.1)	0.560
- Other	2 (1.1)	1 (1.1)	1 (1.1)	1.000
Cardiovascular	15 (8.3)	8 (8.9)	7 (7.8)	0.787
- Coronary calcification	11 (6.1)	6 (6.7)	5 (5.6)	0.756
- Aortic dilation (3–5 cm)	3 (1.7)	1 (1.1)	2 (2.2)	0.560
- Pericardial effusion	1 (0.6)	1 (1.1)	0 (0)	0.316
Thyroid	8 (4.4)	4 (4.4)	4 (4.4)	1.000
- Nodules > 1 cm	5 (2.8)	3 (3.3)	2 (2.2)	0.650
- Diffuse enlargement	3 (1.7)	1 (1.1)	2 (2.2)	0.560
Breast	4 (2.2)	2 (2.2)	2 (2.2)	1.000
- BI-RADS 3 lesions	3 (1.7)	2 (2.2)	1 (1.1)	0.560
- BI-RADS 4 lesions	1 (0.6)	0 (0)	1 (1.1)	0.316
Upper abdominal	3 (1.7)	1 (1.1)	2 (2.2)	0.560
Osseous	1 (0.6)	1 (1.1)	0 (0)	0.316
By Clinical Significance
Low (no follow-up needed)	35 (19.4)	18 (20.0)	17 (18.9)	0.851
Moderate (follow-up recommended)	13 (7.2)	7 (7.8)	6 (6.7)	0.773
High (urgent evaluation needed)	6 (3.3)	3 (3.3)	3 (3.3)	1.000
Clinically Actionable Findings
Suspicious lung nodule	2 (1.1)	1 (1.1)	1 (1.1)	1.000
Thyroid cancer (confirmed)	1 (0.6)	1 (1.1)	0 (0)	0.316
Breast cancer (confirmed)	1 (0.6)	0 (0)	1 (1.1)	0.316
Aortic aneurysm > 5 cm	1 (0.6)	0 (0)	1 (1.1)	0.316
Mediastinal lymphadenopathy	1 (0.6)	1 (1.1)	0 (0)	0.316

Table 5: Spectrum and clinical significance of incidental findings. Abbreviations: SD-CT, standard-dose CT; ULD-CT, ultra-low-dose CT; BI-RADS, Breast Imaging Reporting and Data System. Note: P-values calculated using chi-square test or Fisher's exact test for independent samples.

Predictor	Odds Ratio (95% CI)	P-value
Foreign Body Characteristics
Size < 10 mm (vs ≥10 mm)	3.42 (1.28–9.14)	0.014
Density ≤ 100 HU (vs >100 HU)	2.87 (1.09–7.56)	0.033
Cervical location (vs thoracic)	0.68 (0.24–1.92)	0.467
Patient Factors
BMI > 30 kg/m² (vs ≤30)	2.15 (0.79–5.86)	0.134
Age (per 10 years)	1.08 (0.82–1.42)	0.583
Image Quality Metrics
SNR < 10 (vs ≥10)	4.68 (1.73–12.66)	0.002
Motion artifact present	3.21 (0.94–10.96)	0.063
Protocol
ULD-CT (vs SD-CT)	1.42 (0.48–4.21)	0.527
Interaction Terms
ULD-CT × Low density FB	2.94 (0.87–9.93)	0.083
ULD-CT × BMI >30	1.89 (0.42–8.51)	0.407

Table 6: Multivariable analysis of factors associated with missed or indeterminate foreign bodies. Model performance: C-statistic = 0.84 (95% CI:0.76-0.92) Abbreviations: CI, confidence

interval; HU, Hounsfield units; BMI, body mass index; SNR, signal-to-noise ratio; ULD-CT,

ultra-low-dose CT; SD-CT, standard-dose CT; FB, foreign body.

Supplementary Figure 1: Consolidated standards of reporting trials (CONSORT) flow diagram. The diagram shows patient enrollment, randomization, allocation to standard-dose CT (SD-CT) or ultra-low-dose CT (ULD-CT) protocols, reference standard evaluation with endoscopy/surgery, and final analysis. All 180 randomized patients completed their assigned protocol and were included in the intention-to-treat analysis. EFB = esophageal foreign body; FBP = filtered back projection; ASiR-V = adaptive statistical iterative reconstruction-V; DLIR = deep learning image reconstruction. Please click here to download this file.

Supplementary File 1: De-identified raw data of this study. Please click here to download this file.

Discussion

This prospective randomized trial demonstrates that ULD-CT with DLIR achieves diagnostic accuracy that is non-inferior to that of SD-CT for EFB detection while reducing radiation exposure by 63%. The maintained detection of clinically significant incidental findings suggests that dose reduction need not compromise this important secondary benefit, supporting the potential for ULD-CT as a primary imaging strategy for suspected EFBs across adult and pediatric populations. The success of dramatic dose reduction while maintaining diagnostic accuracy can be attributed to several synergistic factors. First, the inherent high contrast between most foreign bodies and the surrounding soft tissues provides a favorable signal-to-noise environment that tolerates increased image noise^10,28. Even low-density materials such as fishbones (typically 80–150 HU) maintain sufficient contrast against esophageal soft tissue (40–60 HU) for reliable detection²⁹. Second, current evidence on iterative and DLIR reconstruction indicates that dose reduction is most reliable when the diagnostic task is high-contrast and when edge preservation is maintained. ASiR-V and other hybrid iterative methods lower quantum noise but may produce texture changes at high blending strengths, whereas DLIR algorithms use learned image priors to suppress noise while maintaining spatial resolution^17,18,30. Phantom, chest, abdominal, and lung-nodule studies have consistently shown lower image noise and preserved or improved diagnostic confidence with DLIR compared with filtered back projection or conventional iterative reconstruction, even under low-dose or ultra-low-dose conditions^{8,19,21,22,31}. Third, the combination of low-kV (100 kV) imaging with DLIR provides additional benefits through an increased photoelectric effect, enhancing the contrast for both foreign bodies and iodinated contrast when used³².

Critical protocol steps that influence diagnostic success include (1) appropriate tube voltage selection (100 kV optimizes contrast for soft tissue and bone while enabling dose reduction); (2) a dual-reconstruction approach combining iterative reconstruction with deep learning (maximizes noise reduction while preserving edge definition); (3) thin-slice reconstruction with overlap (0.625/0.5 mm enables multiplanar reformations critical for small foreign body detection); (4) a standardized scan range from C3 to the gastroesophageal junction (ensures complete esophageal coverage); and (5) consistent window settings for evaluation (soft tissue: 350/40 HU). Although objective image quality metrics showed the expected degradation with dose reduction—including a 37% decrease in SNR and 58% increase in noise—these changes did not translate to clinically significant diagnostic impairment. This apparent paradox highlights the important distinction between image quality and diagnostic efficacy²⁴. The maintenance of diagnostic acceptability in 94.4% of ULD cases suggests that current clinical protocols may utilize higher doses than necessary for this specific indication. The slight reduction in performance for low-density foreign bodies (sensitivity 93.8% vs. 96.9%, difference -3.1%, 95% CI: -13.0% to 6.8%) warrants consideration but must be contextualized within broader risk-benefit calculations. Using BEIR VII models with appropriate consideration of model uncertainties, this protocol would prevent approximately 47 theoretical cancers per 100,000 pediatric examinations—a population benefit that likely outweighs the small risk of missing low-density foreign bodies that often pass spontaneously^14,33. Furthermore, the availability of immediate endoscopy for equivocal cases provides a safety net that differs from other CT applications, where missed findings might go undetected.

The findings of this study extend and contextualize recent evidence on foreign body imaging and complications. Liu et al. analyzed 275 EFB cases retrospectively, reporting a material spectrum and complication rates that align with the prospective cohort: 42% were bone fragments, and 28% were food impactions, with 3.6% experiencing perforation⁶. Topaloglu et al. highlighted the distinct diagnostic and management challenges in patients with intellectual disabilities, including delayed recognition and a higher likelihood of non-food or sharp-edged foreign bodies requiring surgical management⁴. The randomized imaging-first approach described here confirms that ULD-CT can reliably detect diverse materials while reducing radiation exposure, but patients with impaired communication or persistent symptoms should retain a low threshold for endoscopic confirmation. Shishido et al. characterized species-level differences in fishbone foreign bodies, demonstrating that certain species produce lower-density, more fragmented bones that challenge radiologic detection³⁴. This supports the subgroup finding of reduced sensitivity for low-density foreign bodies (≤100 HU) with ULD-CT (93.8% vs. 96.9% for SD-CT, interaction P = 0.083). When clinical suspicion for sharp fishbones remains high, and CT findings are equivocal—particularly for species known to fragment or appear radiolucent—expedited endoscopic assessment should be favored despite excellent overall ULD-CT performance. Corbisiero et al. reported traumatic vocal fold paralysis from fishbone impaction, highlighting serious laryngeal complications³⁵. This protocol's complete cervical coverage (C3-GEJ) and high sensitivity for cervical foreign bodies (OR: 0.68 for cervical vs. thoracic location, P = 0.467) provide reassurance for detecting foreign bodies in this critical region, where complications can be devastating. The integration of these findings suggests a nuanced triage approach: ULD-CT serves effectively as first-line imaging across most presentations, but for high clinical suspicion of low-density fishbones, unusual ingested objects, or special patient populations with limited symptom reporting, clinicians should maintain lower thresholds for endoscopic evaluation.

The 30% prevalence of incidental findings in this study cohort aligns with previous reports in chest CT, although the rate of clinically actionable findings (3.3%) was lower than that of some series, reporting up to 7%^23,36,37. This difference likely reflects the younger population (median age 44 years) presenting with acute symptoms rather than the older populations typically studied for lung cancer screening or cardiovascular evaluation. The detection of two malignancies (thyroid and breast cancer), which received curative treatment, demonstrates meaningful secondary benefits beyond the primary indication. Recent studies have emphasized the importance of systematic evaluation and appropriate follow-up of incidental findings, with some authors calculating that the life-years gained from incidental cancer detection may offset theoretical radiation risks in appropriately selected populations^38,39. The similar distribution and clinical significance of incidental findings across protocols suggest that dose reduction need not compromise this secondary benefit, although this study was not powered to detect differences in rare but critical findings, such as pulmonary emboli or aortic dissection.

The excellent diagnostic accuracy achieved in pediatric patients by both protocols (100% sensitivity and specificity) supports aggressive dose reduction in this radiosensitive population. Children's smaller body habitus and lower tissue attenuation facilitate imaging at ULDs while maintaining diagnostic quality⁴⁰. The absolute dose achieved in this pediatric cohort (mean 1.37 mSv) approaches that of conventional chest radiography while providing cross-sectional imaging advantages. This is particularly relevant given the high frequency of radiolucent foreign bodies in children, including plastic toys and food items poorly visualized on radiography⁴¹. No motion artifacts requiring repeat imaging occurred in the pediatric cohort, although sedation was used in eight children <5 years old (27% of pediatric patients). Implementation of weight-based protocols with DLIR could potentially achieve even lower doses in the smallest patients, although this study maintained a consistent percentage dose reduction across weight categories to facilitate comparison.

Future applications of this technology should explore several promising directions. Artificial intelligence-assisted interpretation could further enhance diagnostic confidence, particularly for challenging low-density foreign bodies, by providing automated detection and characterization algorithms trained on large datasets. Multi-vendor validation studies are essential to confirm generalizability across different CT platforms and reconstruction algorithms, as this single-vendor study may not fully represent performance on alternative systems. Clinical workflow integration requires the development of standardized protocols, quality assurance metrics, and training programs to ensure consistent implementation across diverse practice settings. Additionally, further dose optimization using photon-counting detector technology may enable sub-millisievert examinations while maintaining or improving image quality.

Beyond radiation dose reduction, considerations of efficiency and usability support ULD-CT adoption. The examination time was identical between protocols (mean 12 seconds scan time, 8 minutes total room time), ensuring no workflow disruption. The image reconstruction time averaged 35 seconds for ULD-CT (deep learning reconstruction) versus 8 seconds for SD-CT (filtered back projection), a minimal delay acceptable in emergency settings. Reproducibility was excellent across scanners, technologists, and readers within the participating institution, with operator-level variability in diagnostic accuracy <2%, suggesting protocol robustness across routine emergency workflows. Resource utilization beyond radiation dose was comparable, with no additional equipment, contrast, or personnel requirements for ULD-CT implementation.

Several important limitations warrant consideration. First, this single-center, single-vendor study utilizing one manufacturer's CT scanners with proprietary deep learning reconstruction may not generalize to other platforms, although similar algorithms are now available from multiple manufacturers. Institutions seeking to implement ULD-CT protocols on non-evaluated systems should conduct local validation studies. For replication using alternative systems, the tube current and reconstruction algorithm strength should be adjusted to achieve a target noise index of 15–20 and SNR > 8. Second, exclusion of patients with morbid obesity (BMI > 40 kg/m²), severe cardiorespiratory instability, pregnancy, prior esophageal surgery or known stricture, substantial metal artifacts, and delayed presentations limits applicability to these important patient populations. The BMI threshold was a prespecified image-quality safeguard because photon starvation and noise amplification under ULD conditions may disproportionately affect low-density foreign body detection; therefore, results should not be extrapolated to severe obesity without local validation or automatic escalation to higher-dose protocols. Third, although endoscopic or surgical reference-standard evaluation was applied to all randomized participants and 30-day follow-up was performed for negative cases, endoscopy may miss small or recently passed foreign bodies, so verification bias cannot be completely eliminated. Fourth, the study was not powered to detect differences in rare but important complications such as perforation (occurred in four patients, all correctly identified by both protocols in their respective arms) or vascular injury (zero cases), which would require substantially larger sample sizes. Fifth, although we demonstrated a similar distribution of incidental findings between protocols, longer-term follow-up would be needed to assess the clinical impact and outcomes of the detected findings. Finally, the 6-h inclusion window may limit generalizability to delayed presentations, particularly fishbones, which commonly present >24 h after ingestion. Future studies should evaluate ULD-CT performance in delayed presentations and in special patient populations where symptom reporting, cooperation, or body habitus may alter imaging characteristics.

Ultra-low-dose CT with DLIR achieves diagnostic accuracy non-inferior to that of SD-CT for EFB detection while reducing radiation exposure by 63%. The protocol maintains excellent performance across adult and pediatric populations and preserves the ability to detect clinically significant incidental findings. These results show potential for ULD-CT as a first-line imaging modality for suspected EFBs pending multicenter validation across diverse CT platforms and patient populations. Future research should explore further dose optimization using photon-counting detector technology, evaluate artificial intelligence-assisted interpretation to enhance diagnostic confidence, and assess long-term outcomes in specific high-risk populations excluded from this trial.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Advantage Workstation (AW) 4.7	GE Healthcare	https://www.gehealthcare.com/en-us/products/imaging-applications/advanced-visualization-applications/advantage-workstation	Imaging workstation
ASiR-V (Adaptive Statistical Iterative Reconstruction-V)	GE Healthcare	N/A	Adaptive statistical iterative reconstruction
AW Server 3.2 Ext 4.0	GE Healthcare	https://www.gehealthcare.com/en-us/products/advanced-visualization-platforms/aw-server	CT system software
Centricity Universal Viewer 6.0	GE Healthcare	https://www.gehealthcare.com/en-us/products/software/enterprise-imaging/centricity-universal-viewer	PACS viewer
MedCalc version 20.0 (MedCalc Software; RRID:SCR_015044)	MedCalc Software	Version 20.0	Medical statistics
Organ Dose Modulation	GE Healthcare	N/A	Organ-based tube current modulation
pROC package for R	CRAN Repository	Version 1.18.4	ROC curve analysis
R version 4.3.2 (R Foundation for Statistical Computing)	R Foundation	RRID:SCR_001905	Statistical computing
Revolution CT (256-slice)	GE Healthcare	https://www.gehealthcare.com/en-us/products/computed-tomography/revolution	CT scanner
Smart mA	GE Healthcare	N/A	Automatic tube current modulation
Standard kernel (soft tissue)	GE Healthcare	N/A	Reconstruction kernel
TrueFidelity DLIR (Deep Learning Image Reconstruction), Medium strength	GE Healthcare	https://www.gehealthcare.com/en-us/products/computed-tomography/applications/true-fidelity	Deep learning image reconstruction