Quantifying the impact of slice thickness on cardiovascular risk stratification in lung cancer screening: a multi-center “RESCUE” study

Introduction

The identification of subclinical atherosclerosis is pivotal for preventing coronary heart disease (CHD), particularly in young adults where traditional risk predictions are often insufficient. Through opportunistic screening of patients undergoing non-gated chest computed tomography (CT) for general health checkups or atypical chest discomfort to rule out non-cardiac pulmonary conditions, we encounter a concerning diagnostic limitation: a substantial proportion of patients who subsequently develop or are confirmed to have obstructive coronary disease presented with a coronary artery calcium (CAC) score of zero on standard thick-slice chest CT scans. This recurring observation of “invisible” risk on standard imaging highlights a critical blind spot in current stratification strategies, potentially delaying essential preventive therapy for high-risk individuals.

With the advancement of artificial intelligence (AI), opportunistic CAC screening on routine chest CTs has emerged as a scalable solution to identify at-risk patients without additional radiation or cost (1-5). However, the standard reconstruction parameters of these scans, which typically utilize a 5.0 mm slice thickness, may be the root cause of the false negatives observed in our cohort. Thick slices suffer from the partial volume effect (PVE), where small, high-density calcium deposits, commonly found in early-stage disease, are averaged with surrounding lower-density tissue and effectively obscured (6). While AI algorithms show high agreement with experts (7-15), they cannot recover calcifications that are physically averaged out by the image reconstruction process.

To address this gap and quantify the magnitude of this “hidden” risk, we conducted a multi-dataset study comparing paired thin- and thick-slice reconstructions. We introduce the concept of the “RESCUE” phenomenon (risk underestimation with thick slices), defined as the reclassification of a patient from CAC =0 on standard thick slices (5.0 mm) to CAC >0 on available thin slices (1.0–2.0 mm). Utilizing a validated deep learning algorithm, we analyzed data of 2,914 patients from four independent cohorts: an internal cohort of early-onset CHD patients (in which the clinical problem was first identified), the National Lung Screening Trial (NLST) dataset, the Stanford AIMI Coronary Calcium and chest CT’s (COCA) dataset for reference validation, and the TotalSegmentator dataset for robustness. This study aims to demonstrate that employing routine thin-slice reconstructions can “rescue” these patients from false-negative assessments, significantly improving the sensitivity of opportunistic screening. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0318/rc).

Methods

Study design

This retrospective, multi dataset diagnostic accuracy study evaluated the impact of CT slice thickness on AI based CAC scoring.

Datasets and validation framework

To assess the “RESCUE” phenomenon, we analyzed imaging data from four independent datasets totaling 2,914 patients, employing a four-stage validation framework:

• Reference standard validation (COCA dataset): to establish AI accuracy against human expert annotations, we utilized the COCA dataset, comprising 651 cases from multiple U.S. centers (444 gated cardiac CT and 207 non-gated chest CT scans). This cohort provided expert manual annotations of coronary calcium, serving as the reference standard for direct AI vs. expert comparison.
• Paired-protocol validation (internal dataset): to quantify the “RESCUE” rate using within-patient comparisons, we evaluated an internal cohort of 766 patients from Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China. To construct this cohort, we queried an invasive coronary angiography database to identify individuals meeting the age criteria for early-onset CAD (men <55 years, women <65 years) who also had a routine non-gated chest CT performed within the three months prior to their angiographic procedure. These preceding baseline scans were acquired strictly for non-cardiac indications, primarily for annual health checkups or the initial triage of atypical chest discomfort. Such cases provide an ideal real-world model for opportunistic CAC screening. Furthermore, we specifically selected this early-onset population because their early-stage, small-volume calcifications are highly susceptible to partial volume averaging, making them the most likely to benefit from thin-slice re-evaluation. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the Ethics Committee of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. SH9H-2019-T160-2). Informed consent was obtained from all patients in the internal cohort.
• External validation (NLST dataset): to confirm generalizability, we identified a systematic, consecutive cohort of participants from the NLST, a multi-center lung cancer screening study conducted across 33 U.S. medical centers (2,3). To avoid selection bias and appropriately match the scale of our primary internal validation cohort, we iteratively downloaded cases from The Cancer Imaging Archive (TCIA) in strictly ascending patient ID order until reaching 857 qualifying participants. Five cases were subsequently excluded due to incomplete or artifact-degraded imaging data, yielding 852 valid cases for final analysis. These 852 paired scans represent a consecutive sample of eligible reconstructions acquired up to our data freeze, ensuring an unbiased representation.
• Real world robustness (TotalSegmentator dataset): to test performance across diverse scanner vendors and clinical indications, we included 645 cases from the TotalSegmentator dataset (16). This cohort originated from University Hospital Basel and 8 collaborating European institutions, covering diverse clinical indications (trauma, oncology, infectious disease) and multiple manufacturers (Siemens, GE, Philips, Toshiba).

CT imaging protocols and reconstruction strategy

For the paired-thickness cohorts (internal and NLST), each patient contributed two series reconstructed from the identical non-gated CT volume acquisition. The thick-slice series utilized a 5.0 mm slice thickness, representing the prevalent clinical standard. The paired thin-slice series used 1.0, 1.5, or 2.0 mm thicknesses, depending on the cohort. Specifically, the internal dataset was divided into two resolution comparison groups: 420 patients with 1.0 vs. 5.0 mm reconstructions, and 346 patients with 1.5 vs. 5.0 mm reconstructions. The NLST external cohort utilized 2.0 vs. 5.0 mm reconstructions. The TotalSegmentator dataset consisted of predominantly 1.5 mm slice thickness reconstructions.

All thin- and thick-slice images (including 5.0 mm) were independently reconstructed from the same raw projection data after a single helical CT scan, using the scanner’s native software. No artificial averaging or post-processing was performed. For the NLST dataset, both paired series were downloaded directly from the official repository database as individual, unmanipulated acquisitions. This paired design allowed within-patient comparison, isolating the effect of slice thickness from physiologic confounding.

Reconstruction kernel selection varied by cohort. The internal cohort employed a medium-smooth kernel (I40f) for thin slices and a medium-sharp kernel (I50f) for thick slices, reflecting routine clinical practice where thin slices are generated for pulmonary nodule detection without additional radiation. The NLST cohort used matched medium-sharp kernels (B50f) for both thin and thick slices, providing a cleanly controlled comparison free from kernel confounding.

AI based CAC scoring algorithm

We used a deep learning algorithm developed for opportunistic screening on non-gated CT scans (14), building on prior validated approaches (7-12). The same model version was applied across all datasets. The algorithm performs three steps: (I) automated cardiac region segmentation; (II) identification of calcified lesions using a 130 Hounsfield unit (HU) threshold, with a minimum lesion area filter (≥1 mm², approx. 3 pixels) to suppress isolated high noise voxels; (III) computation of Agatston scores using standard density weighted factors (17). The algorithm processes entire CT volumes automatically without cardiac gating.

Risk stratification followed standard clinical categories: zero calcium (Agatston 0), minimal to mild [1–99], moderate [100–399], and severe (≥400) (1,18-21). We used these traditional cutoffs to maintain consistency with historical data, while acknowledging the SCCT Coronary Artery Calcium Data and Reporting System (CAC-DRS) guidelines (22).

Statistical analysis

Continuous variables were summarized using means ± standard deviations or medians [interquartile ranges] depending on distribution. For paired comparisons, Pearson correlation coefficients quantified linear associations between thin slice and thick slice Agatston scores, with Spearman rank correlation as sensitivity analysis. Risk category agreement was assessed using exact agreement rates, within one category agreement, and linear weighted kappa statistics.

The primary outcome was the false negative reclassification rate, defined as the proportion of patients with zero calcium on thick slice CT (Agatston =0) but detectable calcium on thin slice CT (Agatston >0). Secondary analyses included: (I) correlation and agreement metrics for paired Agatston scores; (II) stratified distributions of thin slice scores among reclassified patients; (III) net risk category reclassification rates; and (IV) a multivariable logistic regression model constructed among patients with a 5.0-mm CAC score of 0 to identify independent clinical predictors of hidden calcification (RESCUE). All statistical analyses were performed using Python version 3.13 (Python Software Foundation, Wilmington, DE, USA) and standard scientific libraries (pandas, NumPy, SciPy, scikit-learn, statsmodels). Two-sided P<0.05 was considered statistically significant. Institutional Review Board approval was obtained from the participating institutions, and written informed consent was obtained.

Results

Baseline demographic and clinical characteristics are presented in Table 1.

Assessment of algorithm performance (COCA dataset)

To ensure that subsequent findings reflected physical image properties rather than algorithmic errors, we first validated the AI tool against expert human annotations. In the non-gated chest CT subset (n=207), the AI algorithm demonstrated strong correlation with the reference standard (Pearson r=0.956, P<0.001) (Figure 1A, Table 2). The Bland-Altman analysis showed a mean bias of −107.3 Agatston units (Figure 1B), indicating a slight tendency for the AI to underestimate calcium burden compared to experts in challenging non-gated scans, yet 94.7% of patients were correctly classified within ±1 risk category (Figure 1C). These results support the algorithm’s suitability for comparative slice thickness analysis.

Figure 1 AI performance validation against expert reference standard (COCA dataset, n=207). (A) Scatter plot showing the correlation between AI derived Agatston scores (y-axis) and expert reference standard annotations (x-axis). The Pearson correlation coefficient (r=0.956, P<0.001) indicates strong agreement. The red line represents the linear regression fit, while the grey dashed line represents the line of identity. (B) Bland-Altman plot assessing the agreement between AI and the expert reference standard. The mean difference (bias) was −107.35, with 95% limits of agreement ranging from −873.06 to 658.37. (C) Confusion matrix illustrating the concordance of cardiovascular risk categorization (zero, mild, moderate, severe) between AI and the expert reference standard. The heatmap intensity reflects the number of patients in each cell. AI, artificial intelligence; COCA, Stanford AIMI Coronary Calcium and chest CT’s; LoA, limits of agreement; SD, standard deviation.

The “RESCUE” phenomenon in paired cohorts (internal dataset)

To validate the clinical relevance of the “RESCUE” phenomenon, we analyzed the prevalence of obstructive coronary artery disease (CAD) in symptomatic patients who presented with an Agatston score of 0 on standard 5.0 mm scans. Notably, 31% of these patients had obstructive disease (>50% stenosis), and a subset of these cases had calcifications that were only detectable on thin slice reconstructions (Figure 2A). Comparison of clinical characteristics between the “true negative” group (score 0 on both scans) and the “hidden risk” group (score 0 on 5 mm but >0 on thin slices) revealed no significant differences in traditional risk factors (Figure 2B), indicating that clinical assessment alone cannot reliably identify these high-risk individuals.

Figure 2 Clinical validation of the “hidden risk” (internal cohort, n=766). (A) Pie chart showing the prevalence of obstructive CAD (>50% stenosis) among symptomatic patients who had an Agatston score of 0 on standard 5.0 mm scans. Notably, 31% of these patients had obstructive disease, and a subset had calcifications detectable only on thin slices. (B) Comparison of clinical characteristics between the “true negative” group (score 0 on both 5 mm and thin slices) and the “hidden risk” group (score 0 on 5 mm but >0 on thin slices). There were no significant differences (NS, P>0.05) in traditional risk factors, indicating that clinical assessment alone cannot identify these high-risk individuals. CAD, coronary artery disease.

We observed a graded relationship between slice thickness and calcium detection sensitivity. In the high-resolution comparison (1.0 vs. 5.0 mm, n=420), the overall false negative reclassification rate (RESCUE rate) was 19.0% (80/420). However, among the specific subgroup of patients with an initial thick slice Agatston score of 0, the rescue rate was substantially higher at 34.8% (80/230), as illustrated in Figure 3A. Notably, among these rescued patients, 96% (77/80) fell into the minimal to mild risk category (Agatston 1–99) on thin slices, identifying them as candidates for early preventive measures rather than high risk interventions (Table 3).

Figure 3 False negative reclassification and risk reclassification. (A) Bar chart illustrating the rescue rate (false negative reclassification) among patients with an initial Agatston score of 0 on standard 5.0 mm scans. The rates were significantly higher in this subgroup compared to the overall population: 34.8% (80/230) for 1.0 mm, 21.3% (33/155) for 1.5 mm, and 33.8% (91/269) for 2.0 mm (NLST). In the total population (including those with initial score >0), the reclassification rates were 19.0%, 9.5%, and 10.2%, respectively. Data for 1.0 and 1.5 mm are derived from the internal cohort (non-gated chest CT, n=766), while data for 2.0 mm are derived from the NLST cohort (valid paired cases n=852). 95% confidence intervals were computed using the Wilson score method. (B) Sankey diagram depicting the flow of cardiovascular risk categorization from standard 5.0 mm scans (left) to thinner reconstructions (right). This diagram pools data from both internal and NLST cohorts (total paired n=1,618) to visualize the overall trend of risk reclassification. CT, computed tomography; NLST, National Lung Screening Trial.

In the moderate resolution comparison (1.5 vs. 5.0 mm, n=346), the overall RESCUE rate was 9.5% (33/346), corresponding to 21.3% (33/155) of the zero calcium population. Consistent with the 1.0 mm group, the majority (91%) of these reclassified patients had mild calcium scores [1–99]. For patients with calcium visible on both scans, thin slice reconstructions yielded significantly higher Agatston scores, yet maintained strong linear correlation with standard thick slices (r=0.929 for 1.0 mm; r=0.968 for 1.5 mm). Despite the systematic increase in scores on thin slices, risk categorization showed substantial agreement (weighted kappa: 0.705 for 1.0 mm and 0.770 for 1.5 mm) (Table 2).

In multivariable logistic regression analysis of patients presenting with an initial 5.0-mm CAC score of 0 (Table 4), male sex (OR: 3.69; 95% CI: 1.26–10.82; P=0.018) and a history of hypertension (OR: 2.22; 95% CI: 1.08–4.55; P=0.029) emerged as significant independent predictors for experiencing the “RESCUE” phenomenon, identifying subgroups at particularly high risk for thick-slice false negatives.

External validation of the “RESCUE” effect (NLST dataset)

To confirm the generalizability of our findings, we evaluated the “RESCUE” phenomenon in the multi-center NLST external validation cohort (n=852). Comparing 2.0 vs. 5.0 mm reconstructions, 87 patients (10.2%) were reclassified from zero to positive calcium, representing 33.8% of those with an initial zero score (Figure 3A). This 10.2% overall rate closely mirrors the 9.5% rate observed in the internal 1.5 mm group, supporting the generalizability of the directional effect. Notably, 98.9% (86/87) of these reclassified patients had mild calcium scores [1–99], suggesting that thin slice analysis preferentially recovers low burden calcification that can be obscured by partial volume averaging. The overall flow of risk reclassification across both the internal and NLST cohorts is illustrated in Figure 3B, and detailed cross-tabulations of these risk transition pathways from standard to thinner slices are provided in Table 5. Quantitative analysis in the NLST cohort showed excellent correlation between paired scans (r=0.972), with a high-risk category agreement rate of 76.6% and a weighted kappa of 0.816 (Table 2).

Real world robustness (TotalSegmentator dataset)

Finally, to verify the robustness of our approach across diverse imaging platforms, we analyzed the heterogeneous TotalSegmentator cohort (n=645), which included scans from four major manufacturers (Siemens, GE, Philips, Toshiba) acquired for diverse indications. Here, the algorithm quantified calcification across a wide dynamic range, from minimal lesions to extensive burden, as shown in the log scale histogram of positive Agatston scores (Figure 4A). Furthermore, the AI provided consistent risk stratification across this diverse dataset (Figure 4B), identifying patients across all risk categories (zero, mild, moderate, severe). This supports the viability of the tool, as well as the strategy of opportunistic screening, across the variability inherent in real world clinical practice.

Figure 4 Real world robustness in a heterogeneous dataset (TotalSegmentator, n=645). (A) Histogram (log scale) showing the distribution of positive Agatston scores detected by the AI. The algorithm quantified calcification across a wide dynamic range, from minimal lesions to extensive burden. (B) Distribution of cardiovascular risk categories in the TotalSegmentator cohort. The AI provided consistent risk stratification in this diverse dataset acquired from multiple scanner types, supporting its generalizability to real world clinical settings. AI, artificial intelligence.

Discussion

In this multi-center study of 2,914 patients, we evaluated the impact of slice thickness on cardiovascular risk stratification in opportunistic screening. As CAC assessment on routine non-gated chest CT becomes increasingly common, interpreting a “zero calcium” score requires consideration of protocol limitations. We observed that a subset of patients classified as CAC =0 on standard 5.0 mm reconstructions demonstrated detectable calcification on paired thin slice images from the same acquisition, which we term the “RESCUE” phenomenon. This reclassification was observed in 10.2% of patients in the external NLST cohort (2.0 mm) and 19.0% in the internal high-resolution cohort (1.0 mm). These findings align with and significantly expand upon recent smaller-scale observations that thinner reconstructions improve detection sensitivity (23). Importantly, this trend was consistent across our large, multi-dataset cohorts, supporting a technical rather than population specific etiology.

Algorithm reliability and generalizability

Before interpreting the clinical significance of these findings, it is essential to establish that the additional calcifications detected on thin slices represent true anatomical structures rather than image noise or algorithmic artifacts. We utilized a deep learning algorithm specifically optimized for coronary calcification, distinct from general purpose organ segmentation models (24). Validation against expert annotations in the COCA dataset yielded a strong correlation (r=0.956), confirming that the algorithm performs comparably to human readers and aligning with prior validations in non-gated settings (8). Furthermore, the algorithm demonstrated robust performance across diverse scanner manufacturers in the TotalSegmentator dataset, consistent with findings by Xu et al. regarding cross vendor generalizability (25). This consistency suggests that the “RESCUE” phenomenon is not an algorithmic hallucination but a reproducible finding. Although higher image noise in thin slice scans is a known challenge in low dose CT (13), the high specificity observed in our validation cohorts indicates that the deep learning model effectively distinguishes true high-density calcification from stochastic noise.

Mechanism of the “RESCUE” phenomenon

With the algorithm’s reliability established, the observed reclassification is primarily attributable to the PVE (6,26). When calcified plaques are smaller than the voxel dimensions of thick slice images, their high-density signal is averaged with surrounding soft tissue, reducing attenuation below the standard 130 HU threshold (17). Consequently, threshold-based scoring on thick slices may systematically miss small lesions, consistent with prior clinical data by Willemink et al. showing that PVEs systematically lower Agatston scores (27). Our data reflect this resolution dependent sensitivity: 1.0 mm reconstructions identified additional calcium in 19.0% of cases, compared to approximately 10% for 1.5 and 2.0 mm protocols. This confirms that the “RESCUE” phenomenon is driven by the physical recovery of spatial resolution, allowing the detection of small plaques that are mathematically invisible on standard 5 mm reconstructions.

Although recent studies have explored the utility of thin-slice CT for CAC scoring (23,25,28-34), these investigations have largely been limited by small sample sizes or single-center designs. For instance, Shin et al. (31), Lee et al. (33), and Ki et al. (34) validated thin-slice protocols in cohorts of fewer than 30 patients, while Lin et al. (32) evaluated AI scoring in 76 patients. Early work by Kim et al. (35) utilizing retrospective reconstruction techniques was limited to 128 participants. Furthermore, Groen et al. (36) demonstrated the clinical applicability of non-gated ordinal scoring in 140 patients, but did not address the specific benefits of AI-driven quantitative rescue. Similarly, while Suh et al. (37) validated automated scoring in a multi-center cohort of 452 patients, their analysis primarily focused on agreement with manual readers rather than the specific impact of slice thickness on false-negative rates and risk reclassification. Some studies specifically focused on verifying deep learning-based image denoising or style transfer techniques rather than the intrinsic value of native thin-slice data for risk reclassification (25,29). In contrast, our study capitalizes on the largest multi-center cohort to date (n=2,914) to demonstrate that the “RESCUE” phenomenon is a robust, physically-driven effect. Unlike approaches relying on AI-generated “simulated” standard images (30), we validate that native thin-slice reconstruction parameters alone can recover a substantial proportion of false-negative risk without requiring additional complex post-processing. While Yin et al. (23) reported similar trends, our study strictly differentiates itself by validating this in a nearly 30-fold larger population (2,914 vs. 112 patients) and, crucially, linking this technical reclassification to the presence of obstructive CAD.

Clinical implications: the “power of one” and targeted prevention

The most significant clinical implication of our findings lies in the reclassification of patients from the “zero calcium” category to the “minimal/mild” risk category (Agatston 1–99). Current guidelines often use CAC =0 as a “gatekeeper” to defer statin therapy, relying on its high negative predictive value (the “Power of Zero”) (22,38). However, our study suggests that for up to 19% of patients, this “zero” is a technical artifact of thick slice imaging rather than a biological reality. The transition from Agatston 0 to >0 (the “Power of One”) fundamentally shifts the prevention strategy from reassurance to active risk factor modification. While long-term prospective outcome data for “rescued” patients are still needed, our finding that 31% of symptomatic patients with CAC =0 on 5 mm scans (but potentially CAC >0 on thin slices) actually harbor obstructive coronary disease (Figure 2A) highlights the immediate clinical relevance. Given that thin-slice reconstructions (≤2 mm) are already the standard of care for modern lung cancer screening and routine chest CT, utilizing this data for AI-based CAC assessment represents a “zero-additional-cost” and “zero-additional-dose” strategy to reduce false-negative risk stratification. This approach allows clinicians to provide more sensitive and personalized preventive care without any changes to existing scan workflows.

The most immediate impact is seen in patients with borderline to intermediate 10-year ASCVD risk (5% to <20%), where the 2018 AHA/ACC Cholesterol Guidelines recommend CAC scoring as a “tie-breaker” (19). In this “gray zone”, a score of CAC =0 aids in improved risk refinement, whereas a score of CAC 1–99 supports informing shared clinical decision-making regarding early preventive interventions. A false negative “zero” on a thick slice scan could therefore mask genuine, early-stage CAD.

Furthermore, younger patients and women often present with lower overall calcium burdens that are particularly susceptible to partial volume averaging. In our study, the “RESCUE” phenomenon was predominantly observed in the mild risk range [1–99]. For women younger than 60, the presence of any CAC (>0) typically places them above the 75th percentile for their age and race, indicating a high lifetime risk trajectory that warrants aggressive lifestyle modification (18,39). Similarly, for younger adults (<40 years), detecting early calcification shifts the focus from low 10-year risk to elevated lifetime risk, justifying earlier intervention.

This refinement in detection sensitivity is also crucial for interpreting epidemiological data. Recent large-scale studies on opportunistic screening (4,5) likely included individuals with occult calcification in their “CAC =0” reference groups, potentially underestimating the true event rate of a strictly negative scan. Prioritizing thin slice analysis, particularly in lung cancer screening programs where such data are routinely available, aligns with the 2025 AHA statement on opportunistic screening and emerging perspectives on asymptomatic risk assessment, offering a “zero additional dose” strategy to maximize cardiovascular prevention (40,41).

Limitations

Our study has limitations. First, the study design is retrospective. Second, while the TotalSegmentator cohort provides real world diversity, it lacks paired thick slice reconstructions for direct false negative reclassification analysis. Third, expert reference standard annotations were not available at scale for the internal and NLST cohorts; paired analyses therefore relied on within scan higher resolution comparisons. Fourth, reconstruction kernel heterogeneity exists across datasets, although the consistent findings across cohorts with different kernels suggest the effect is primarily driven by slice thickness. Fifth, we lack long term clinical outcome data to validate the prognostic significance of mild calcifications detected only on thin slices. Sixth, vessel specific analysis of false negative patterns was not performed. Finally, it is important to recognize that the increased sensitivity of thin-slice scans comes with an inherent trade-off. Thinner reconstructions inherently display higher image noise, which can occasionally elevate the risk of false-positive calcium detection if excessive noise artifacts exceed the 130 HU threshold. While our algorithm utilizes minimum area filters to mitigate this concern, clinicians must interpret solitary, small high-density voxels on thin slices with caution.

Conclusions

Standard 5 mm reconstructions can miss low-burden coronary calcification in up to about 19% of patients. Routinely available thin-slice reconstructions (≤2 mm) with AI detect most of these missed lesions and reclassify roughly 10–19% of patients, mainly to Agatston 1–99, requiring no extra radiation or workflow changes. Thin-slice AI analysis improves early detection and helps guide preventive care, ensuring that a zero score reliably indicates the absence of calcification.

Acknowledgments

We would like to thank the research teams and participants of the Stanford AIMI (Center for Artificial Intelligence in Medicine and Imaging) “Coronary Calcium and chest CT’s” (COCA) dataset, the National Lung Screening Trial (NLST), and the TotalSegmentator dataset (University Hospital Basel) for the public availability of their data. We also acknowledge the clinical and technical support provided by the Department of Cardiology, Shanghai Ninth People’s Hospital.

Footnote

Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0318/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0318/dss

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0318/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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the Ethics Committee of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. SH9H-2019-T160-2). Informed consent was obtained from all patients in the internal cohort.

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. Hecht HS, Cronin P, Blaha MJ, Budoff MJ, Kazerooni EA, Narula J, Yankelevitz D, Abbara S. 2016 SCCT/STR guidelines for coronary artery calcium scoring of noncontrast noncardiac chest CT scans: A report of the Society of Cardiovascular Computed Tomography and Society of Thoracic Radiology. J Cardiovasc Comput Tomogr 2017;11:74-84. [Crossref] [PubMed]
2. National Lung Screening Trial Research Team. Aberle DR, Adams AM, Berg CD, Black WC, Clapp JD, Fagerstrom RM, Gareen IF, Gatsonis C, Marcus PM, Sicks JD. Reduced lung-cancer mortality with low-dose computed tomographic screening. N Engl J Med 2011;365:395-409.
3. Aberle DR, Berg CD, Black WC, Church TR, Fagerstrom RM, Galen B, Gareen IF, Gatsonis C, Goldin J, Gohagan JK, Hillman B, Jaffe C, Kramer BS, Lynch D, Marcus PM, Schnall M, Sullivan DC, Sullivan D, Zylak CJ. The National Lung Screening Trial: overview and study design. Radiology 2011;258:243-53.
4. Peng AW, Dudum R, Jain SS, Maron DJ, Patel BN, Khandwala N, Eng D, Chaudhari AS, Sandhu AT, Rodriguez F. Association of Coronary Artery Calcium Detected by Routine Ungated CT Imaging With Cardiovascular Outcomes. J Am Coll Cardiol 2023;82:1192-202. [Crossref] [PubMed]
5. Malik RF, Sun KJ, Azadi JR, Lau BD, Whelton S, Arbab-Zadeh A, Wilson RF, Johnson PT. Opportunistic Screening for Coronary Artery Disease: An Untapped Population Health Resource. J Am Coll Radiol 2024;21:880-9. [Crossref] [PubMed]
6. Mühlenbruch G, Thomas C, Wildberger JE, Koos R, Das M, Hohl C, Katoh M, Günther RW, Mahnken AH. Effect of varying slice thickness on coronary calcium scoring with multislice computed tomography in vitro and in vivo. Invest Radiol 2005;40:695-9. [Crossref] [PubMed]
7. Martin SS, van Assen M, Rapaka S, Hudson HT Jr, Fischer AM, Varga-Szemes A, Sahbaee P, Schwemmer C, Gulsun MA, Cimen S, Sharma P, Vogl TJ, Schoepf UJ. Evaluation of a Deep Learning-Based Automated CT Coronary Artery Calcium Scoring Algorithm. JACC Cardiovasc Imaging 2020;13:524-6. [Crossref] [PubMed]
8. van Velzen SGM, Lessmann N, Velthuis BK, Bank IEM, van den Bongard DHJG, Leiner T, de Jong PA, Veldhuis WB, Correa A, Terry JG, Carr JJ, Viergever MA, Verkooijen HM, Išgum I. Deep Learning for Automatic Calcium Scoring in CT: Validation Using Multiple Cardiac CT and Chest CT Protocols. Radiology 2020;295:66-79. [Crossref] [PubMed]
9. Eng D, Chute C, Khandwala N, Rajpurkar P, Long J, Shleifer S, et al. Automated coronary calcium scoring using deep learning with multicenter external validation. NPJ Digit Med 2021;4:88. [Crossref] [PubMed]
10. Isgum I, Prokop M, Niemeijer M, Viergever MA, van Ginneken B. Automatic coronary calcium scoring in low-dose chest computed tomography. IEEE Trans Med Imaging 2012;31:2322-34. [Crossref] [PubMed]
11. Lessmann N, van Ginneken B, Zreik M, de Jong PA, de Vos BD, Viergever MA, Isgum I. Automatic Calcium Scoring in Low-Dose Chest CT Using Deep Neural Networks With Dilated Convolutions. IEEE Trans Med Imaging 2018;37:615-25. [Crossref] [PubMed]
12. Zeleznik R, Foldyna B, Eslami P, Weiss J, Alexander I, Taron J, et al. Deep convolutional neural networks to predict cardiovascular risk from computed tomography. Nat Commun 2021;12:715. [Crossref] [PubMed]
13. Kim S, Park EA, Ahn C, Jeong B, Lee YS, Lee W, Kim JH. Performance of fully automated deep-learning-based coronary artery calcium scoring in ECG-gated calcium CT and non-gated low-dose chest CT. Eur Radiol 2025;35:7084-95. [Crossref] [PubMed]
14. Hagopian R, Strebel T, Bernatz S, Myers GA, Offerman E, Zuniga E, Kim CY, Ng AT, Iwaz JA, Nürnberg L, Singh SP, Carey EP, Kim MJ, Schaefer RS, Yu J, Gentili A, Aerts HJWL. AI Opportunistic Coronary Calcium Screening at Veterans Affairs Hospitals. NEJM AI 2025; [Crossref]
15. Mineo E, Assuncao-Jr AN, Grego da Silva CF, Liberato G, Dantas-Jr RN, Graves CV, Gutierrez MA, Nomura CH. Automated Coronary Artery Calcium Scoring Using Deep Learning: Validation Across Diverse Chest CT Protocols. Acad Radiol 2025;32:7177-86. [Crossref] [PubMed]
16. Wasserthal J, Breit HC, Meyer MT, Pradella M, Hinck D, Sauter AW, Heye T, Boll DT, Cyriac J, Yang S, Bach M, Segeroth M. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiol Artif Intell 2023;5:e230024. [Crossref] [PubMed]
17. Agatston AS, Janowitz WR, Hildner FJ, Zusmer NR, Viamonte M Jr, Detrano R. Quantification of coronary artery calcium using ultrafast computed tomography. J Am Coll Cardiol 1990;15:827-32. [Crossref] [PubMed]
18. Arnett DK, Blumenthal RS, Albert MA, Buroker AB, Goldberger ZD, Hahn EJ, Himmelfarb CD, Khera A, Lloyd-Jones D, McEvoy JW, Michos ED, Miedema MD, Muñoz D, Smith SC Jr, Virani SS, Williams KA Sr, Yeboah J, Ziaeian B. 2019 ACC/AHA Guideline on the Primary Prevention of Cardiovascular Disease: A Report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines. J Am Coll Cardiol 2019;74:e177-232. [Crossref] [PubMed]
19. Grundy SM, Stone NJ, Bailey AL, Beam C, Birtcher KK, Blumenthal RS, et al. 2018 AHA/ACC/AACVPR/AAPA/ABC/ACPM/ADA/AGS/APhA/ASPC/NLA/PCNA Guideline on the Management of Blood Cholesterol: A Report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines. Circulation 2019;139:e1082-143. [Crossref] [PubMed]
20. Greenland P, Blaha MJ, Budoff MJ, Erbel R, Watson KE. Coronary Calcium Score and Cardiovascular Risk. J Am Coll Cardiol 2018;72:434-47. [Crossref] [PubMed]
21. Mahabadi AA, Möhlenkamp S, Lehmann N, Kälsch H, Dykun I, Pundt N, Moebus S, Jöckel KH, Erbel RHeinz Nixdorf Recall Study Investigators. CAC Score Improves Coronary and CV Risk Assessment Above Statin Indication by ESC and AHA/ACC Primary Prevention Guidelines. JACC Cardiovasc Imaging 2017;10:143-53. [Crossref] [PubMed]
22. Hecht HS, Blaha MJ, Kazerooni EA, Cury RC, Budoff M, Leipsic J, Shaw L. CAC-DRS: Coronary Artery Calcium Data and Reporting System. An expert consensus document of the Society of Cardiovascular Computed Tomography (SCCT). J Cardiovasc Comput Tomogr 2018;12:185-91.
23. Yin K, Chen W, Qin G, Liang J, Bao X, Yu H, Li H, Xu J, Chen X, Wang Y, Savage RH, Schoepf UJ, Mu D, Zhang B. Performance assessment of an artificial intelligence-based coronary artery calcium scoring algorithm in non-gated chest CT scans of different slice thickness. Quant Imaging Med Surg 2024;14:5708-20. [Crossref] [PubMed]
24. McCollough CH, Ulzheimer S, Halliburton SS, Shanneik K, White RD, Kalender WA. Coronary artery calcium: a multi-institutional, multimanufacturer international standard for quantification at cardiac CT. Radiology 2007;243:527-38. [Crossref] [PubMed]
25. Xu C, Guo H, Xu M, Duan M, Wang M, Liu P, Luo X, Jin Z, Liu H, Wang Y. Automatic coronary artery calcium scoring on routine chest computed tomography (CT): comparison of a deep learning algorithm and a dedicated calcium scoring CT. Quant Imaging Med Surg 2022;12:2684-95. [Crossref] [PubMed]
26. Šprem J, de Vos BD, Lessmann N, van Hamersvelt RW, Greuter MJW, de Jong PA, Leiner T, Viergever MA, Išgum I. Coronary calcium scoring with partial volume correction in anthropomorphic thorax phantom and screening chest CT images. PLoS One 2018;13:e0209318. [Crossref] [PubMed]
27. Willemink MJ, Vliegenthart R, Takx RA, Leiner T, Budde RP, Bleys RL, Das M, Wildberger JE, Prokop M, Buls N, de Mey J, Schilham AM, de Jong PA. Coronary artery calcification scoring with state-of-the-art CT scanners from different vendors has substantial effect on risk classification. Radiology 2014;273:695-702. [Crossref] [PubMed]
28. Kang HW, Ahn WJ, Jeong JH, Suh YJ, Yang DH, Choi H, Hwang SH, Yong HS, Oh YW, Kang EY, Kim C. Evaluation of fully automated commercial software for Agatston calcium scoring on non-ECG-gated low-dose chest CT with different slice thickness. Eur Radiol 2023;33:1973-81. [Crossref] [PubMed]
29. Choi H, Park EA, Ahn C, Kim JH, Lee W, Jeong B. Performance of 1-mm non-gated low-dose chest computed tomography using deep learning-based noise reduction for coronary artery calcium scoring. Eur Radiol 2023;33:3839-47. [Crossref] [PubMed]
30. Kim C, Hong S, Choi H, Yoo WS, Kim JY, Chang S, Park CH, Hong SJ, Yang DH, Yong HS, van Assen M, De Cecco CN, Suh YJ. Impact of Deep Learning-Based Image Conversion on Fully Automated Coronary Artery Calcium Scoring Using Thin-Slice, Sharp-Kernel, Non-Gated, Low-Dose Chest CT Scans: A Multi-Center Study. Korean J Radiol 2025;26:759-70. [Crossref] [PubMed]
31. Shin JM, Kim TH, Kim JY, Park CH. Coronary artery calcium scoring on non-gated, non-contrast chest computed tomography (CT) using wide-detector, high-pitch and fast gantry rotation: comparison with dedicated calcium scoring CT. J Thorac Dis 2020;12:5783-93. [Crossref] [PubMed]
32. Lin Y, Lin G, Peng MT, Kuo CT, Wan YL, Cherng WJ. The Role of Artificial Intelligence in Coronary Calcium Scoring in Standard Cardiac Computed Tomography and Chest Computed Tomography With Different Reconstruction Kernels. J Thorac Imaging 2024;39:111-8. [Crossref] [PubMed]
33. Lee SY, Kim TH, Han K, Shin JM, Kim JY, Kim D, Park CH. Feasibility of Coronary Artery Calcium Scoring on Dual-Energy Chest Computed Tomography: A Prospective Comparison with Electrocardiogram-Gated Calcium Score Computed Tomography. J Clin Med 2021;10:653. [Crossref] [PubMed]
34. Ki SJ, Park CH, Han K, Shin JM, Kim JY, Kim TH. Utility of the 16-cm Axial Volume Scan Technique for Coronary Artery Calcium Scoring on Non-Enhanced Chest CT: A Prospective Pilot Study. Taehan Yongsang Uihakhoe Chi 2021;82:1493-504. [Crossref] [PubMed]
35. Kim SM, Chung MJ, Lee KS, Choe YH, Yi CA, Choe BK. Coronary calcium screening using low-dose lung cancer screening: effectiveness of MDCT with retrospective reconstruction. AJR Am J Roentgenol 2008;190:917-22. [Crossref] [PubMed]
36. Groen RA, Jukema JW, van Dijkman PRM, Timmermans PT, Bax JJ, Lamb HJ, de Graaf MA. Evaluation of Clinical Applicability of Coronary Artery Calcium Assessment on Non-Gated Chest Computed Tomography, Compared With the Classic Agatston Score on Cardiac Computed Tomography. Am J Cardiol 2023;208:92-100. [Crossref] [PubMed]
37. Suh YJ, Kim C, Lee JG, Oh H, Kang H, Kim YH, Yang DH. Fully automatic coronary calcium scoring in non-ECG-gated low-dose chest CT: comparison with ECG-gated cardiac CT. Eur Radiol 2023;33:1254-65. [Crossref] [PubMed]
38. Dzaye O, Dardari ZA, Cainzos-Achirica M, Blankstein R, Agatston AS, Duebgen M, Yeboah J, Szklo M, Budoff MJ, Lima JAC, Blumenthal RS, Nasir K, Blaha MJ. Warranty Period of a Calcium Score of Zero: Comprehensive Analysis From MESA. JACC Cardiovasc Imaging 2021;14:990-1002. [Crossref] [PubMed]
39. Osei AD, Mirbolouk M, Dardari Z, Shea S, Blankstein R, Dzaye O, Nasir K, Blumenthal RS, Blaha MJ. A Simple Approach to the Identification of Guideline-Based Coronary Artery Calcium Score Percentiles (From the Multi-Ethnic Study of Atherosclerosis). Am J Cardiol 2022;179:18-21. [Crossref] [PubMed]
40. Foraker R, Sperling L, Bratzke L, Budoff M, Leppert M, Razavi AC, Rodriguez F, Shapiro MD, Whelton S, Wong ND, Yang EAmerican Heart Association Prevention Science Committee of the Council on Epidemiology and Prevention and Council on Cardiovascular and Stroke Nursing. Council on Clinical Cardiology; and Council on Cardiovascular Radiology and Intervention. Opportunistic Detection of Coronary Artery Calcium on Noncardiac Chest Computed Tomography: An Emerging Tool for Cardiovascular Disease Prevention: A Scientific Statement From the American Heart Association. Circulation 2025;152:e391-401. [Crossref] [PubMed]
41. Kawaguchi YO, Fujimoto S, Nozaki YO, Tomizawa N, Daida H, Minamino T. Current status and future perspective of coronary artery calcium score in asymptomatic individuals. J Cardiol 2025;85:275-82. [Crossref] [PubMed]