Designed as a multicenter prospective observational study, it included patients who underwent health check-ups at 3 health institutions affiliated with the Korea Association of Health Promotion (KAHP) between December 2022 and December 2023. Consecutive recruitment was implemented according to inclusion and exclusion criteria. Inclusion criteria were adults aged 20 years or older who had undergone abdominal CT scans as part of their routine health check-ups. Exclusion criteria were patients with canceled or incomplete CT scans, or those with significant artifacts on CT images. No specific screening for comorbidities such as diabetes or prior malignancies was performed. We included all patients who underwent abdominal CT, regardless of contrast enhancement.

CT scans were performed using Revolution Discovery (GE Healthcare), Somatom Definition AS+, and Somatom Definition Edge (Siemens Healthineers) scanners, using parameters as follows: tube voltage, 120 kVp; effective tube current, 200 reference mAs (Care Dose 4D; Siemens Healthineers) or 100‐400 mA (AutomA or SmartmA; GE Healthcare); field of view, 30‐40 cm; collimation, 0.31‐0.75 mm; and pitch, 0.98‐1.00. Images were reconstructed using the filtered backprojection technique with the soft tissue reconstruction algorithm at a section thickness of 5 mm with no interslice gap. To ensure that variations between scanners did not affect segmentation accuracy or Hounsfield Unit (HU) measurements, we standardized acquisition protocols with the same kVp, reconstruction technique, and slice thickness.

The AI solution (Aid-U, iAID Inc.), a fully automatic deep learning solution for both L3 selection and body composition assessment, was used for the AI algorithm model (Video S1 in Multimedia Appendix 1). The software uses a You Only Look Once v3 (YOLOv3)–based algorithm for automatic L3 inferior endplate level selection and a fully convolution network for segmentation of abdominal muscle area and fat areas, as illustrated in Figure 1 [9]. Selected L3 level CT images were automatically segmented to generate the boundary of total abdominal muscles and measured the abdominal muscle and fat area [10]. Based on predefined HU thresholds, our system then classified the segmented muscle into normal-attenuation muscle area (NAMA, +30 HU to +150 HU), low-attenuation muscle area (LAMA, −29 HU to +29 HU), and inter- and intramuscular adipose tissue (IMAT, −190 HU to −30 HU). The distinction between NAMA, LAMA, and IMAT was made to differentiate healthy muscle from fatty infiltration and to assess myosteatosis. This approach allows a more detailed assessment of muscle quality [11,12]. We defined a skeletal muscle area (SMA) by combining NAMA and LAMA. The BMI-adjusted SMA is defined by SMA/BMI. The visceral fat area (VFAT) and the subcutaneous fat area (SFAT) were demarcated using fat tissue thresholds of −190 HU to −30 HU information.

To ensure seamless compatibility with existing clinical workflows and institutional imaging infrastructures, the AI model (Aid-UTM, iAID Inc.) was directly integrated with the PACS (INFINITT PACS, Infinitt Co.) at the participating health institutions. Processes for the fully automated AI-integrated opportunistic CT screening system were established in a step-by-step manner, following the most recent implementation guidelines, as illustrated in Figure 2 [13].

The first step was pilot implementation. A project kick-off meeting was conducted with the AI engineering team, the clinical team from KABP, and the PACS team to ensure seamless coordination. During this phase, the AI solution servers were integrated with the PACS infrastructure. Retrospective testing was performed using test CT datasets (n=10) obtained from the open data repository of the 2022 Body Morphometry AI Segmentation Challenge. The second step involved limited clinical deployment. CT scans from the first 10 patients enrolled in this study were used to evaluate real-time AI processing in a prospective manner. An internal monitoring tool was deployed alongside this step to track CT data transfer and AI processing performance, ensuring robustness and transparency. The final step was full clinical deployment, enabling the system for long-term use in routine health check-up settings. This phase marked the integration of the AI system into daily clinical workflows for opportunistic CT screening, ensuring scalability and sustained performance.

The technical success rate of the AI system was calculated as the proportion of CT scans successfully processed without requiring manual adjustments. The overall processing time was reported as the mean time interval from the completion of CT scan acquisition to the availability of the AI-generated report. Segmentation accuracy was assessed by comparing AI-generated results with manual annotations by radiologists, with agreement quantified using the Dice similarity coefficient (DSC).

To evaluate age-related and gender-specific differences in body composition metrics, 1-way ANOVA was conducted across age groups. P<.05 was considered statistically significant.

Our study proposed the concept of muscle age, defined as the estimated age at which the participant’s muscle mass corresponds to the average muscle mass of that age group. Muscle age is calculated using data from Kim et al’s study [14], which includes 11,845 healthy Korean adults who underwent abdominal CT scans during routine health check-ups. Participants with known malignancy, chronic diseases, or metabolic disorders were excluded. SMA and SMA/BMI values were stratified by sex and age. Participants aged 35 to 65 years, representing the 5th to 95th percentile of the healthy adult population, were used to generate sex-specific regression models for estimating muscle age. First, we regressed SMA/BMI on a constant and age in a sample of participants aged between 35 and 65 years, representing approximately the 5th and 95th percentiles for age in both sexes. Next, we derived the inverse of the regression equation by solving for age as a function of SMA/BMI. The SE of the estimated age was computed using the delta method, based on the original variance-covariance matrix adjusted for heteroscedasticity. Finally, separate regression analyses were performed for men and women, resulting in distinct muscle age functions for each sex. The statistical analyses were conducted using R (version 4.3.0; R Foundation for Statistical Computing).

This study was approved by the Asan Medical Center institutional review board (approval code IRB number 2022-1049; approval date August 4, 2022). The requirement for patient consent was waived.

A total of 537 consecutive patients (mean age 46.9, SD 13.2 y; 345 males and 192 females) were recruited for the study. During the recruitment period, no participants were excluded. Most patients underwent abdominal CT scans for initial evaluation or follow-up of incidental lesions detected on routine abdominal ultrasonography such as pancreatic cysts, gallbladder polyps, and focal lesions in the liver, kidney, or spleen. Some patients underwent abdominal CT scans due to increased levels of blood tumor markers. The demographic characteristics of the participants, including height, weight, and BMI, are listed in Table 1.

The AI algorithm model was implemented on on-premise servers which were successfully integrated into the PACS in 3 institutions. As illustrated in Figure 3, to establish fully automated AI-integrated opportunistic CT screening system, the following processes were established:

The fully automated AI-integrated system provided detailed body composition metrics, including muscle quality maps and estimated muscle age, as part of the health check-up reports (Figure 4). The integration of the serial analysis function was added for tracking longitudinal changes. In addition, a muscle age calculator was developed so that clinical users can calculate muscle age manually if DICOM metadata does not contain height and weight information (Video S2 in Multimedia Appendix 2; assessed on the web [15]).

The AI system demonstrated robust performance across all evaluated metrics. Among the 537 abdominal CT scans analyzed, the technical success rate was 100%, with no cases requiring manual adjustment or failing due to technical errors.

The mean overall processing time from the completion of CT scan acquisition to the generation of AI results was 4.12 (SD 0.15) seconds per scan, confirming the system’s real-time applicability.

Segmentation accuracy of the AI algorithm, as assessed by the DSC, was 0.974 (SD 0.011), indicating excellent agreement with manual radiologist annotations.

Results of opportunistic screening for body composition are presented in Table 2. In terms of sarcopenia screening, age-related changes in muscle mass and quality were evident in both male and female participants, as NAMA and SMA, indicators of healthy muscle and total muscle, respectively, consistently declined with age. In contrast, LAMA and IMAT, which reflect fat infiltration and reduced muscle quality, showed significant increases across age groups (Figure 5). These trends align with the natural progression of sarcopenia and myosteatosis, emphasizing the impact of aging on muscle health.

Regarding obesity screening, VFAT increases with age in both genders, highlighting the trend of age-related visceral fat accumulation, a known risk factor for metabolic disorders. SFAT, however, exhibited differing trends between males and females. In males, SFAT decreased in older age groups, while in females, it remained relatively stable across all ages, suggesting a relative preservation of subcutaneous fat. These differences may point to sex-specific mechanisms in fat storage and distribution.

This prospective multicenter study evaluated the implementation of a fully automated AI-integrated system of body composition assessment for opportunistic CT screening during routine health check-ups. The AI system demonstrated robust performance, with a 100% technical success rate, an average processing time of 4.12 seconds per scan, and high segmentation accuracy (DSC: mean 0.9740.011). The AI system provided comprehensive body composition metrics, including muscle quality maps and estimated muscle age, seamlessly integrated into health check-up reports. These findings validate the reliability, efficiency, and scalability of the AI system for opportunistic sarcopenia screening.

The study revealed significant age-related changes in muscle quality and fat distribution, consistent with the progression of sarcopenia and myosteatosis [11,12,16]. In both males and females, NAMA and SMA declined with advancing age, reflecting a reduction in muscle mass and quality. In contrast, LAMA and IMA, markers of muscle fat infiltration, increased significantly with age. These results underscore the capability of the AI system to detect and quantify meaningful variations in body composition, aligning with physiological expectations and clinical relevance [1].

We used BMI-adjusted skeletal muscle index (SMA/BMI) in our study. Although height-adjusted skeletal muscle index (SMA/height²) is widely used in Western populations, previous research in a large Korean cohort demonstrated that the BMI-adjusted index (SMA/BMI) better reflects age-related muscle decline. Therefore, we adopted SMA/BMI in this study [14].

Notably, our study proposed the concept of muscle age, which provides an individualized estimate of physiological muscle health by projecting the participant’s muscle mass onto the age group with a similar average muscle mass. This metric allows for direct comparison between an individual’s muscle mass and population norms, enabling more nuanced assessment beyond traditional metrics such as chronological age or raw muscle mass, like the concept of biological age [17].

There have been numerous retrospective studies which have assessed AI systems for body composition analysis in terms of segmentation accuracy, technical adequacy, and prognostic utility [11,12,18-20]. Retrospective studies, though valuable, often lack the operational insights necessary to understand the challenges and benefits of integrating AI into real-time clinical workflows. To date, only a few prospective studies have explored the use of AI-based systems in specific clinical contexts, such as routine abdominal CT scans for pancreas cancer follow-up or lung cancer screening programs [21,22].

To the best of our knowledge, this is the first prospective multicenter study to implement a fully automated AI-integrated system for opportunistic screening of body composition in health check-up institutions. By evaluating the system’s performance across diverse institutions, this study bridges the gap between research applications and practical clinical implementation, highlighting the readiness of AI technologies to transition from the research setting to routine health care practice [13]. The successful integration of the AI system into clinical workflows demonstrates its potential use in routine health check-ups by automating time-intensive tasks and report generation. Furthermore, the system’s ability to detect early signs of sarcopenia and myosteatosis offers significant clinical value, as timely interventions can prevent progression to frailty and associated complications [6]. Despite these promising results, integrating AI into institutional PACS presented several challenges, including ensuring seamless data transfer, maintaining compliance with DICOM standards, and obtaining clinical team acceptance. A phased deployment approach and early collaboration between AI engineers, radiologists, and PACS teams were essential to overcoming these barriers. For real-world implementation, we are preparing regulatory approval as a software as a medical device and ensuring compliance with regulations such as the Health Insurance Portability and Accountability Act (HIPAA) and the General Data Protection Regulation (GDPR).

This study has several limitations. First, the study population consisted of participants undergoing routine health check-ups, which may limit the generalizability of the findings to other clinical settings, such as acute care or cancer follow-up. Second, while the AI system demonstrated high segmentation accuracy and efficiency, the study relied on standardized imaging protocols across the participating institutions. Variations in CT acquisition protocols or differences in scanner types at other facilities might affect the system’s performance, necessitating additional validation in diverse imaging environments. Finally, the study focused on technical performance, without incorporating long-term patient outcomes or clinical decision-making for patient care. Despite these limitations, this study represents an important step toward integrating fully automated AI systems into routine clinical practice, offering a foundation for future research to address these gaps.

This study demonstrated the successful implementation of a fully automated AI-integrated system for body composition assessment using opportunistic CT screening during routine health check-ups, achieving high accuracy, efficiency, and scalability. Significant age-related changes in muscle quality and fat distribution were identified, highlighting the system’s clinical utility in detecting sarcopenia and myosteatosis.