Real-World Implementation of Artificial Intelligence: Fully Automated Coronary Artery Calcium Scoring

INTRODUCTION

Coronary artery calcium (CAC) scoring derived from non-contrast ECG-gated cardiac CT, most commonly expressed as the Agatston score, is a well-established tool for atherosclerotic cardiovascular disease risk stratification and is a marker of coronary artery disease strongly endorsed by major preventive cardiology guidelines [1]. CAC provides incremental prognostic value beyond traditional cardiovascular risk factors, particularly in individuals at intermediate risk [1].

Despite its well-established clinical utility, routine CAC assessments remain challenging in daily practice. Manual CAC measurement is labor-intensive, requires specialized expertise, and is subject to inter-reader and protocol-related variability [2,3]. These limitations may result in reporting delays and inconsistencies in high-volume clinical settings. Although semi-automated tools exist, substantial manual input is still required, which limits scalability and workflow efficiency [2].

To overcome the limitations of routine manual CAC assessment, CAC automation has been developed and has progressively evolved from earlier feature-based or atlas-only approaches to contemporary deep-learning-based frameworks, most commonly leveraging convolutional neural networks for calcium detection and anatomical segmentation [4]. Deep learning–based CAC quantification enables rapid and reproducible measurements with increasing vendor-agnostic performance. Multiple studies and meta-analyses have demonstrated that artificial intelligence (AI)-based CAC scoring systems can reliably reproduce the Agatston scores and categorical risk grading across diverse clinical settings, reflecting the technical maturity of contemporary AI-based approaches [5,6]. Emerging evidence suggests that AI-derived plaque metrics, including vessel-specific calcium burden, may enhance cardiovascular risk prediction [7,8].

Recently, the application of AI-based automatic CAC scoring has been extended to non-ECG-gated chest CT, demonstrating reliable measurements [9]. However, variability among AI models and limited real-world validation across both cardiac and non-cardiac CT have constrained their widespread adoption, prompting a growing emphasis on reliability, end-to-end automation, vessel-specific quantification, and seamless integration into routine clinical workflows, beyond algorithmic performance alone [10,11].

To address these challenges, we implemented an AI-based CAC scoring system in routine clinical practice and shared our real-world experience focusing on its practical benefits and limitations. In this editorial, we focus on ECG-gated noncontrast cardiac CT, the reference standard for CAC scoring, rather than non-ECG-gated chest CT, given the current lack of standardized and robust quantitative approaches for AI-based CAC assessment on chest CT [9].

INSTITUTIONAL EXPERIENCE WITH THE IMPLEMENTATION OF AUTOMATED CAC SCORING

Our institution is a university-affiliated tertiary hospital and a regional cardiovascular and cerebrovascular center that operates an integrated health-screening program. Approximately 170 electrocardiogram (ECG)-gated cardiac CT examinations are performed monthly, most of which involve CAC scoring. Cardiac CT examinations are performed using multiple CT platforms from different vendors, reflecting a heterogeneous imaging environment typical of real-world practice. While all cardiac CT examinations are interpreted by a single radiologist, CAC analysis has traditionally been performed by radiographers using semi-automated software that requires manual region-of-interest placement and vessel labeling. This workflow is time-consuming and operator-dependent, and frequently leads to reanalysis and workflow disruption.

In Korea, AI-based CAC scoring systems received regulatory approval in July 2021. AI-based CAC scoring has been applied to all ECG-gated non-contrast cardiac CT examinations at our institution since April 2025. The implemented system (AVIEW CAC version 1.1, Coreline Soft., Seoul, South Korea) combined an atlas-guided approach with deep learning–based semantic segmentation, enabling the anatomically constrained identification of coronary calcifications while minimizing the misclassification of noncoronary calcium [4]. CAC results were automatically generated and uploaded to the picture archiving and communication system (PACS; INFINITT G7, INFINITT Healthcare, Seoul, South Korea) alongside the original CT data, providing total and vessel-specific metrics, including Agatston, volume, and mass scores, as well as percentile-based risk stratification according to the patient’s age and sex, derived from population-based reference cohorts (Fig. 1) [12]. Grayscale soft presentation state-based overlays are available, but are not routinely used in our institutional workflow.

Fig. 1. Example of AI-based CAC scoring integrated into the routine PACS workflow. A: Screenshot of the PACS showing an AI-based CAC scoring report uploaded alongside the corresponding ECG-gated noncontrast cardiac CT images, enabling simultaneous review during routine interpretation. B: Axial ECG-gated non-contrast cardiac CT image demonstrating calcified plaque in the LAD (arrowhead). C: Magnified view cropped from the AI-based CAC scoring report, showing automated detection and vessel-specific labeling of LAD calcification overlaid in purple. D: AI-generated CAC summary report demonstrating a total Agatston score of 99, with vessel-specific CAC metrics and age- and sex-adjusted, percentile-based risk stratification corresponding to the 50th–75th percentile. AI = artificial intelligence, CAC = coronary artery calcium, PACS = picture archiving and communication system, ECG = electrocardiogram, LAD = left anterior descending artery.

Automated CAC scoring markedly improved workflow efficiency, with AI-generated CAC reports typically becoming available on PACS within 10 minutes of CT image upload. The initial implementation phase was accompanied by transmission-related delays and intermittent AI report delivery issues; however, these were resolved with system stabilization, allowing reliable integration into routine clinical workflows.

In our institutional audit, we observed a high level of agreement between AI-based CAC scoring and manual reference standards for both continuous scores and categorical grading, particularly for clinically relevant thresholds (Fig. 1). Nevertheless, occasional misclassification was noted between the left main and left anterior descending arteries in selected patients. False-positive and false-negative detections also persisted, most commonly related to non-coronary calcifications and motion artifacts. These limitations were more pronounced in patients with a high CAC burden or structural coronary abnormalities such as coronary dilatation, occlusion, or anatomic variants (Fig. 2).

Fig. 2. False-negative detection and vessel misclassification in a patient with coronary artery ectasia and an intraluminal thrombus. A: Axial non-contrast ECG-gated cardiac CT image demonstrating multifocal calcifications within the LCx (arrowhead). B: On the AI-based CAC output, some calcified lesions were incorrectly classified as LAD calcifications and highlighted in purple (arrow), whereas other hyperattenuating foci within the LCx were identified as ≥130-Hounsfield unit pixels but were not classified as CAC, appearing only as pink markings (arrowhead). C: Contrast-enhanced CT image demonstrating marked ectasia of the LCx with intraluminal thrombus. Arrowhead indicates true CAC within the intramural thrombus. ECG = electrocardiogram, LCx = left circumflex artery, AI = artificial intelligence, CAC = coronary artery calcium, LAD = left anterior descending artery.

Following its implementation, AI-based CAC scoring has replaced radiographer-generated preliminary reports, thereby reducing repetitive manual tasks, unnecessary reanalysis, and operator-dependent variability. These workflow optimizations allow both radiographers and radiologists to focus on higher-level clinical and interpretive work in high-volume clinical environments.

DISCUSSION

Prior studies have emphasized that successful workflow improvement with AI requires robust integration beyond algorithmic speed [13,14]. Our experience supports this concept, as the rapid generation of AI-based CAC reports was achievable only after stable PACS integration was established.

Multiple studies, including recent meta-analyses, have demonstrated consistently high agreement between AI-based CAC scoring and manual reference standards for both continuous and categorical grading in diverse settings, reflecting the technical maturity of contemporary AI-based approaches [14,15]. Our institutional experience was concordant, particularly in identifying clinically critical thresholds such as CAC zero and CAC ≥100, which directly influence statin initiation under current guidelines [1].

Despite these favorable characteristics, AI-based CAC scoring has several limitations. Previous studies have reported high agreement for vessel-level CAC quantification [5,16]; however, our observation of occasional misclassification between the left main and left anterior descending arteries highlights an area where further algorithmic refinement is warranted. Given the additional prognostic value of vessel-specific CAC assessment [17], accurate territorial classification remains clinically important.

False-positive and false-negative detections have been consistently reported in the literature and are most commonly related to non-coronary calcifications, motion artifacts, excessive CAC burden, and complex coronary anatomy [14,18]. These findings are concordant with our institutional experience and underscore that even with technically mature AI systems, radiologist oversight remains essential in routine clinical practice.

Further improvements in algorithmic accuracy are warranted, particularly in patients with extensive calcification or complex coronary anatomy [14]. Equally important is seamless PACS integration with transparent visualization and editable AI outputs, which enables radiologists to efficiently verify and, when necessary, correct CAC results within routine workflows. Although AI-based CAC scoring has reached practical maturity for routine clinical use, the final responsibility for the accuracy of the reported CAC scores remains with radiologists [10,11].

In conclusion, AI-based CAC scoring improves workflow efficiency and diagnostic consistency in routine practice, supporting its suitability for use in high-volume cardiovascular imaging environments. The key points are summarized as follows:

• • AI-based CAC scoring enables rapid and reliable CAC assessment, substantially reducing manual workload and interoperator variability in routine practice.

  • Despite its technical maturity, radiologist oversight remains essential, particularly in patients with a high CAC burden or complex coronary anatomy, where false-positive or false-negative findings may still occur.

  • With thoughtful clinical implementation and seamless PACS integration, AI-based CAC scoring represents a mature and valuable tool for contemporary high-workload cardiovascular imaging.