AI Solution for Identifying Actionable Abnormalities on Chest Radiographs: A Retrospective Case-Control Study

Abstract

Purpose

To evaluate the diagnostic performance of a commercial artificial intelligence (AI) solution in detecting actionable abnormalities (AAs) on chest radiographs (CRs) and to compare its performance with that of thoracic radiologists.

Materials and Methods

Among 986,016 CRs performed at a single academic hospital between 2016 and 2021, 194 CRs were retrospectively identified as actionable based on critical value reports. Age- and sex-matched 388 normal CRs were selected as controls. A commercial AI solution capable of detecting 10 thoracic abnormalities was retrospectively applied to all CRs. Three thoracic radiologists independently reviewed the same dataset. Diagnostic performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive rate, precision, F1 score, detection yield, and false-referral rate (FRR). Interobserver agreement and agreement between the AI and the radiologists were also assessed.

Results

The AI solution demonstrated excellent diagnostic performance, with an AUC of 0.958, a sensitivity of 90.7%, and a specificity of 91.2%. Compared with the radiologists, the AI showed higher sensitivity and F1 score and achieved the highest detection yield (30.2%, p ≤ 0.02). The FRR of the AI (5.8%) was higher than that of one radiologist (1.7%) but comparable to those of the others. The AI showed substantial agreement (κ = 0.704) with the radiologists’ consensus, while interobserver agreement among the radiologists was moderate (κ = 0.45).

Conclusion

The AI solution demonstrated diagnostic performance comparable to that of thoracic radiologists and may assist in detecting AAs when false-positive alerts are appropriately managed.

Graphical Abstract

INTRODUCTION

Chest radiography is a widely accessible and cost-effective first-line imaging tool for thoracic diseases (1). Its use has rapidly increased, placing growing demands on radiologists who must interpret large image volumes amid ongoing workforce shortages (2,3,4). Even with adequate staffing, radiologists cannot identify critical findings until each chest radiograph (CR) is reviewed.

Timely interpretation of CRs is essential for prompt diagnosis and effective patient management. The American College of Radiology (ACR) Actionable Reporting Work Group defined three categories of imaging findings that require additional, nonroutine communication with the referring physician within minutes, hours, or days, depending on their urgency or unexpected nature (5). These are referred to as actionable abnormalities (AAs), which account for approximately 1.5%–32.5% of radiology reports (2,5,6).

Recent advances in convolutional neural networks have facilitated the development of AI tools capable of interpreting CRs. Numerous studies have demonstrated AI’s performance comparable to that of radiologists in detecting key abnormalities and aiding clinical decision-making (7,8,9). These findings suggest that AI can assist in workflow triage, particularly when reporting delays occur due to imaging overload (10,11). However, some evaluations may overestimate AI performance because they rely on internal datasets similar to the model’s training data (12).

We hypothesized that a commercial AI solution could identify CRs with a high likelihood of containing critical AAs for prompt radiologist review. The clinical utility of AI depends not only on its diagnostic accuracy but also on its capacity to recognize a broad spectrum of abnormalities, including device-related findings such as tube malposition. To our knowledge, no previous study has evaluated the performance of a commercial AI solution in detecting the full range of real-world actionable findings as determined by radiologists. Therefore, this study aimed to assess the diagnostic performance of a commercial AI solution in identifying AAs on CRs—specifically those reported as critical by thoracic radiologists—and to compare its performance with that of thoracic radiologists.

MATERIALS AND METHODS

This was a retrospective, single-center study. The study protocol was approved by the Institutional Review Board of Boramae Medical Center (IRB No. 20-2025-69). All data were de-identified, and the requirement for written informed consent was waived.

STUDY POPULATION AND DATASET

Fig. 1 summarizes the data selection process. We retrospectively reviewed reports of 986,016 CRs obtained with posteroanterior or anteroposterior (AP) projections from individuals aged 18 years or older at our institution between January 2016 and December 2021. In our Picture Archiving and Communication System, when a radiologist identifies a finding on a CR that requires a critical value report (CVR)—which necessitates immediate notification of the referring clinician—a text message is automatically sent to the ordering physician upon report finalization, serving as an alert to review the report. CVR-triggered findings include both acute, life-threatening abnormalities (e.g., tension pneumothorax, pneumoperitoneum) and nonemergent but management-altering findings (e.g., new lung nodules, consolidation). Because of the conceptual equivalence between ACR-defined AAs and our institution’s CVR policy, CVR-triggering findings were designated as AAs (5). Therefore, the designation “AA = CVR” in this study represents an operational adaptation of the ACR framework to a real-world reporting workflow and ensures consistency in defining clinically significant abnormalities.

Fig. 1. Flowchart of case selection for CRs included in the case–control dataset.

AA = actionable abnormality, CR = chest radiograph

To select the case group, we included 545 CRs flagged as CVRs. To focus on newly detected AAs and avoid duplication of cases, we excluded CRs with previously identified AAs and those without any AAs (n = 985,471). Among the CRs flagged as CVRs, 351 were excluded because the AAs were subsequently not confirmed either radiologically (e.g., by chest CT) or clinically. Consequently, 194 CRs with radiologically or clinically confirmed AAs were included as the case group. Assuming a prevalence of CRs with actionable findings of one-third based on the results of a previous study (13), 388 normal CRs were included as controls, each confirmed by chest CT performed within 7 days of CR acquisition. Two normal CRs were matched to each abnormal CR based on sex and an age difference of ≤2 years. The control CRs were selected from examinations performed at our institution during the same period and showed no actionable findings. In total, 582 CRs were randomly assigned for independent assessment by a commercial AI solution and three thoracic radiologists. Data were collected on a one-CR-per-patient basis. Patient demographics and imaging characteristics—including sex, age, radiograph projection, and image referral setting (outpatient, inpatient, or emergency room)—were also analyzed.

REFERENCE STANDARDS FOR AAs

Reference standards for AAs were established by consensus between two board-certified radiologists (J.K.N. and K.M.Y., with 19 and 11 years of experience in thoracic imaging, respectively). They reviewed 545 CRs that had previously been reported as containing AA. For each case, the radiologists reviewed the corresponding chest CT or serial CRs, along with relevant clinical information from the medical records, and documented the actionable findings and final diagnoses.

AI SOLUTION FOR CRs

In this study, we used a commercially available AI solution (Lunit INSIGHT CXR3, version 3.1.3.5) designed to detect 10 common abnormalities on CRs: atelectasis, calcification, cardiomegaly, consolidation, fibrosis, mediastinal widening, nodule or mass, pleural effusion, pneumoperitoneum, and pneumothorax. The AI algorithm is based on a 34-layer residual neural network (ResNet-34), which processes the raw pixel data of Digital Imaging and Communications in Medicine files to generate probability maps and abnormality scores (range, 0–100) for each lesion (14).

To evaluate diagnostic performance, the AI model output was binarized using an optimized threshold of 22.31, determined from the receiver operating characteristic (ROC) curve by maximizing the Youden index (Fig. 2) to best differentiate CRs with and without AAs. For binary classification, the maximum score among the 10 abnormality categories was used to classify a CR as positive (with AA) or negative (without AA). When the AI assigned a positive abnormality score but failed to localize the actionable finding, the case was re-evaluated for the relevance of the annotated area. If the annotation corresponded to a secondary finding related to the AA, it was considered a true positive; if it was unrelated, it was considered a false positive.

Fig. 2. ROC curve of the AI solution. The ROC curve illustrates the diagnostic performance of the AI solution in detecting AAs on chest radiographs. The criterion threshold of the AI abnormality score used to classify a radiograph as positive was set at >22.31. At this operating point, the AI achieved a sensitivity of 90.7% and a specificity of 91.2%, with an AUC of 0.958 (p < 0.001).

AA = actionable abnormality, AUC = area under the ROC curve, ROC = receiver operating characteristic

Additionally, we analyzed the distribution of AI-predicted abnormalities in both CRs with AAs and control CRs, as well as the corresponding false-positive classifications. The mean abnormality scores were compared between CRs with AAs and control CRs and summarized as mean ± standard deviation with 95% confidence intervals (CIs).

COMPARISON BETWEEN THE AI SOLUTION AND THORACIC RADIOLOGISTS

To mitigate the uncertainty in positive and negative predictive values (PPV and NPV) arising from prevalence-related bias and to focus on comparing the relative diagnostic performance between the AI solution and human radiologists, three thoracic radiologists evaluated the same case-control dataset as the AI.

Three board-certified thoracic radiologists with 13, 10, and 9 years of experience in thoracic imaging independently evaluated all 582 CRs. They were instructed to identify AAs and disregard findings deemed clinically insignificant.

The diagnostic performance of the AI solution and thoracic radiologists was assessed using the area under the ROC curve (AUC), sensitivity, and specificity. In addition, positive rate, precision, F1 score, detection yield, and false-referral rate (FRR) were analyzed to enable direct comparison of case-detection efficiency between the AI and the radiologists, reflecting their practical reporting tendencies within the fixed case-control dataset. Precision was defined as the proportion of true-positive cases among all positive predictions, distinguishing diagnostic accuracy from reporting tendency. The F1 score was calculated as the harmonic mean of precision and recall, providing a balanced measure that accounts for both false-positive and false-negative results. Positive rate represented the proportion of cases classified as positive regardless of correctness, indicating each reader’s or the AI’s tendency to interpret CRs as positive for AAs. Detection yield referred to the proportion of actionable cases correctly identified by each reader or by the AI model within the dataset. FRR was defined as the proportion of false-positive cases among all cases, reflecting the frequency of unnecessary follow-ups triggered by incorrect positive predictions. When applicable, 95% CIs were calculated using standard statistical methods.

STATISTICAL ANALYSIS

AUCs for the AI solution and each radiologist were compared using DeLong’s test. Sensitivity, specificity, detection yield, and FRR were compared using McNemar’s test. Positive rates between the AI and the radiologists were compared using two-proportion z tests. P-values for precision and F1 scores were estimated using paired bootstrap resampling (10,000 replicates).

To evaluate interobserver agreement, Fleiss’ kappa and the intraclass correlation coefficient (ICC) were calculated among the three radiologists who independently reviewed all CRs. In addition, diagnostic agreement between the AI and the radiologist consensus—defined as agreement by at least two radiologists—was assessed using percent agreement and Cohen’s kappa.

All statistical analyses were performed using SPSS Statistics (version 26.0; IBM Corp., Armonk, NY, USA) and Rex (version 3.6.0; RexSoft Inc., Seoul, South Korea). For all tests, statistical significance was set at p < 0.05.

RESULTS

Patient demographics and imaging characteristics are summarized in Table 1. The proportion of CRs obtained with an AP projection was significantly higher in the AA group than in the control group (p < 0.001). The most common referral setting in both groups was outpatient, and no statistically significant difference was observed between the two groups in terms of referral setting.

Table 1. Patient Demographics and Imaging Characteristics.

		Total Subjects (n = 582)	CR with AA (n =194)	Normal CR (n = 388)	p-Value
Male		372 (63.9)	124	248	1.000
Female		210 (36.1)	70	140	1.000
Age, yrs*		63.0 ± 17.7	63.8 ± 19.6	62.6 ± 16.7	0.463
CR projection
	Postero-anterior	424 (72.9)	119 (61.3)	305 (78.6)	<0.001
	Antero-posterior	158 (27.1)	75 (38.7)	83 (21.4)	<0.001
Referral setting
	Outpatient	345 (59.3)	114 (58.8)	226 (58.2)	0.854
	Inpatient	75 (12.9)	27 (13.9)	43 (11.1)	0.322
	Emergency room	162 (27.8)	53 (27.3)	119 (30.7)	0.405

Data are presented as mean ± standard deviation or number (percentage), unless otherwise indicated.

AA = actionable abnormality, CR = chest radiograph

Table 2 summarizes the AA and final diagnoses in the AA group. Nodules or masses were the most common AA (84 cases, 43.3%), followed by consolidation (40 cases, 20.1%), pneumothorax/hydropneumothorax (36 cases, 18.6%), and mediastinal widening (23 cases, 11.9%). Less frequent findings included tube malposition, pneumomediastinum, and pneumoperitoneum (each 1.0%–1.5%). Final diagnoses were most commonly pneumothorax/hydropneumothorax (36 cases, 18.6%), pneumonia (28 cases, 14.4%), lung cancer (28 cases, 14.4%), and tuberculosis (26 cases, 13.4%), with various rare conditions classified as “other” (9 cases, 4.6%).

Table 2. Actionable Findings and Final Diagnoses.

Actionable findings
	Nodules or masses	84 (43.3)
	Consolidation	40 (20.1)
	Pneumothorax/hydropneumothorax	36 (13.9)
	Mediastinal widening	22 (11.9)
	Tube malposition	3 (1.5)
	Pneumomediastinum	3 (1.5)
	Pneumoperitoneum	2 (1.0)
	Foreign body	1 (0.5)
	Osteolytic lesions	1 (0.5)
	Cement embolism	1 (0.5)
	Pneumatosis intestinalis	1 (0.5)
	Total	194 (100)
Final diagnosis
	Pneumothorax/hydropneumothorax	36 (18.6)
	Pneumonia	28 (14.4)
	Lung cancer	28 (14.4)
	Tuberculosis	26 (13.4)
	Follow-up loss	20 (10.3)
	Aortic aneurysm	17 (8.8)
	Fibrotic sequelae	7 (3.6)
	Atelectasis	5 (2.6)
	Tube malposition	3 (1.5)
	Other malignancy	3 (1.5)
	Gastrointestinal perforation	3 (1.5)
	Nontuberculous mycobacterial lung disease	3 (1.5)
	Summation opacity	2 (1.0)
	Bronchiectasis	2 (1.0)
	Pulmonary edema	2 (1.0)
	Others*	9 (4.6)
	Total	194 (100)

All data are numbers of patients, with percentages in parentheses.

^*Others include bone island, cement embolism, bronchogenic cyst, sarcoidosis, pneumatosis intestinalis, foreign body, hiatal hernia, aspergilloma, and pulmonary infarction.

Fig. 2 illustrates the ROC curve for the AI solution. The AUC was 0.958 (p < 0.001), with a sensitivity of 90.7% and a specificity of 91.2%. The optimal cutoff value for the AI abnormality score was 22.31.

Table 3 summarizes the diagnostic performance of the AI solution and the three thoracic radiologists. The AI achieved the highest AUC (0.958; 95% CI, 0.938–0.977), which was significantly higher than that of all radiologists (p ≤ 0.011). The sensitivity of the AI (90.7%) was higher than that of Radiologists 2 (35.6%, p < 0.001) and 3 (70.1%, p = 0.034) but not significantly different from Radiologist 1 (85.1%, p = 0.09). Specificity did not differ significantly between the AI and Radiologists 1 and 2 but was lower than that of Radiologist 3 (97.4%, p = 0.047). The AI demonstrated a higher positive rate (36.1%) compared with Radiologist 2 (19.6%, p < 0.001) and Radiologist 3 (25.1%, p < 0.001), while the difference with Radiologist 1 (32.5%) was not statistically significant (p = 0.19). The precision and F1 scores of the AI (83.8% and 87.1%, respectively) were comparable to those of Radiologist 1 (87.3% and 86.2%, both p > 0.05) but significantly higher than those of Radiologists 2 and 3 (all p < 0.02). The detection yield of the AI (30.2%) was significantly higher than that of the radiologists (28.4%, 11.9%, and 23.4%, respectively; p < 0.03). The FRR of the AI (5.8%) was significantly higher than that of Radiologist 3 (1.7%, p < 0.001) but not significantly different from Radiologist 1 (4.1%, p = 0.15) or Radiologist 2 (7.7%, p = 0.22).

Table 3. Diagnostic Performance of the AI Solution and Three Radiologists.

	AI	Radiologist 1	p-Value	Radiologist 2	p-Value	Radiologist 3	p-Value
AUC	95.8% (93.8–97.7)	89.4% (86.6–92.0)	0.003	62.0% (58.2–65.7)	<0.001	83.8% (80.4-86.8)	0.011
Sensitivity	90.7% (86.6–94.8)	85.1% (80.0–90.1)	0.090	35.6% (28.8–42.3)	<0.001	70.1% (63.7-76.5)	0.034
Specificity	91.2% (88.4–94.1)	93.8% (91.4–96.2)	0.221	88.4% (85.2–91.6)	0.147	97.4% (95.8-99.0)	0.047
Positive rate	36.1% (32.3–40.1)	32.5% (28.3–36.9)	0.218	19.6% (15.7–24.1)	<0.001	25.1% (20.8–29.8)	<0.001
Precision	83.8% (78.8–88.8)	87.3% (82.6–92.0)	0.267	60.5% (51.6–69.5)	<0.001	93.2% (89.1-97.2)	<0.001
F1 score	87.1%	86.2%	0.420	44.8%	<0.001	80.0%	0.018
Detection yield	30.2% (26.6–34.1)	28.4% (24.4–32.7)	0.020	11.9% (8.8–15.7)	<0.001	23.4% (19.3–27.8)	<0.001
FRR	5.8% (4.2–8.1)	4.1% (2.6–6.4)	0.150	7.7% (5.3–11.1)	0.220	1.7% (0.8–3.6)	<0.001

Values in parentheses indicate 95% confidence intervals.

AUC = area under the receiver operating characteristic curve, FRR = false-referral rate

The interobserver agreement among the three radiologists was moderate, with a Fleiss’ kappa of 0.45 and an ICC of 0.46. When the AI’s binary output was compared with the consensus diagnosis, the percent agreement was 87.3%, and Cohen’s kappa was 0.704, indicating substantial concordance.

Supplementary Table 1 summarizes the distribution of AI-predicted abnormalities and false-positive classifications. Among CRs with AI abnormality scores greater than 22.31, the most frequently predicted abnormalities were nodules (25.7%), consolidation (22.4%), and pneumothorax (13.8%). Abnormality-specific sensitivity was highest for pneumothorax (100%), nodules (90.5%), and consolidation (87.5%) and moderate for mediastinal widening (80.0%). False positives were most frequently observed in cases classified as cardiomegaly (29.4%), fibrosis (14.7%), and pneumoperitoneum (8.8%). These false-positive classifications—particularly for nonspecific findings such as cardiac silhouette enlargement and fibrotic changes—highlight potential areas for model refinement. The AI system generated 210 classifications above the threshold, of which 34 (16.2%) were false positives from the control group. The mean AI abnormality score was 73.7 ± 29.0 (95% CI, 69.6–77.8) in the AA group, which was significantly higher than that in the control group (8.5 ± 13.2; 95% CI, 7.2–9.8; p < 0.001).

Representative examples of the AI’s performance limitations are presented in Fig. 3. As a false-negative case, miliary tuberculosis with diffuse, tiny nodular opacities was correctly identified by the radiologists but missed by the AI, which generated a low abnormality score (<10) without lesion annotation. Conversely, in cases of tube malposition, the AI did not directly recognize the malpositioned Levin tube but produced an elevated abnormality score (42%) due to cardiomegaly and atelectasis. Similarly, in two cases of retracted percutaneous drainage catheters, the AI yielded high abnormality scores because of secondary pleural effusion, despite failing to annotate the catheter malposition itself. These examples highlight the AI’s limitations in detecting subtle disseminated diseases and device-related abnormalities, although secondary findings may indirectly elevate abnormality scores.

Fig. 3. Representative examples demonstrating the limitations of AI performance.

A, B. A case of miliary tuberculosis that was missed by the AI but correctly identified by the radiologists. Numerous tiny nodular opacities are diffusely distributed throughout both lungs (A), yet the AI generated an abnormality score below 10 with no lesion annotation (B).

C, D. A case of a malpositioned Levin tube (arrow) not directly detected by the AI. However, the abnormality score was elevated (42%) because of secondary findings such as cardiomegaly and atelectasis, despite the AI’s failure to annotate the tube malposition (D).

DISCUSSION

DIAGNOSTIC PERFORMANCE AND AGREEMENT

This study evaluated the diagnostic performance of a commercial AI solution in identifying AAs on CRs, achieving an AUC of 95.8%, with a sensitivity of 90.7% and a specificity of 91.2%. The AI outperformed three thoracic radiologists in sensitivity, F1 score, and detection yield while maintaining an acceptable FRR of 5.8%, supporting its potential clinical utility in detecting critical findings. These findings align with previous studies demonstrating that commercial AI solutions can detect multiple chest abnormalities and assist radiologists in interpretation, particularly in high-volume clinical environments (15,16,17).

To assess the consistency of AI decisions relative to human interpretation, we evaluated interobserver agreement among radiologists and between the AI model and the radiologist consensus. The moderate agreement among the three radiologists (Fleiss’ kappa = 0.45; ICC = 0.46) reflects the inherent variability in interpreting subtle or borderline findings on CRs, consistent with previous reports (18,19). In contrast, the AI model demonstrated substantial agreement with the radiologist consensus (kappa = 0.704), suggesting that its outputs are largely aligned with expert interpretation in detecting AAs. This concordance highlights the AI’s potential to serve as a reliable triage tool in clinical practice, particularly in settings with limited radiologist availability or high imaging workloads. Moreover, the strong agreement with expert consensus indicates that AI may help standardize radiographic interpretation by reducing interobserver variability. Radiologists often differ in their interpretive thresholds—some tend to minimize false positives through conservative judgments, while others prioritize sensitivity to avoid missing actionable findings. This suggests that AI can function as a stabilizing adjunct, promoting greater consistency and balance in radiographic interpretation.

The substantial difference in AI abnormality scores between the AA group (73.7%) and the control group (8.5%) demonstrates strong discriminative ability, reinforcing the AI’s potential as a triage tool for workflow optimization. A recent multicenter validation study also confirmed the feasibility of AI for excluding normal CRs, suggesting its dual utility in both flagging urgent findings and dismissing normal examinations (20,21,22). Our study extends previous triage validations to acute care settings, where 27% of actionable cases originated from emergency departments. The higher proportion of AP projections in the AA group (38.7%) compared with the control group (21.4%; p < 0.001) reflects the predominance of bedside radiography in acute settings. Despite technical variability, the AI maintained stable diagnostic performance across projection types, indicating suitability for real-world clinical deployment (23). Our results complement these findings by focusing specifically on AA detection, providing evidence of the AI’s dual capability in both excluding normal cases and flagging critical findings.

AI PERFORMANCE ACROSS ABNORMALITY TYPES

Although the AI solution demonstrated overall high diagnostic performance, its detection of AAs was not uniform across abnormality types. As shown in Supplementary Table 1, the AI successfully detected most AAs associated with nodules, consolidation, and pneumothorax, consistent with previous studies that evaluated AI models for these abnormalities (10,24).

However, the AI occasionally missed device-related abnormalities, such as tube malposition. In these instances, the AI tended to assign elevated abnormality scores because of secondary findings—such as atelectasis or pleural effusion—rather than identifying the primary abnormality. Although such indirect cues may partially compensate for unrecognized findings, their clinical interpretability remains limited. Further refinement of AI algorithms to explicitly recognize medical devices and line malposition is warranted, as these represent visually subtle yet clinically urgent AAs (25). Moreover, our study found that the AI system occasionally failed to detect subtle presentations such as miliary tuberculosis, echoing the results of a recent prospective, multicenter AI validation study for tuberculosis detection (26). These observations underscore the importance of recognizing the diagnostic boundaries of AI and emphasize the continued necessity of radiologist oversight.

FALSE POSITIVES AND CLINICAL LIMITATIONS

Despite achieving high sensitivity, false positives—particularly those involving cardiomegaly and fibrosis—remain a significant challenge. ROC analysis using an optimal cutoff value of 22% identified these as the most common false-positive classifications, consistent with findings from previous evaluations of AI-assisted double reading (27). The FRR of 5.8% compares favorably with earlier automated triage studies, although it highlights the persistent challenge of balancing sensitivity and specificity in clinical AI applications (12).

The issue of false positives extends beyond statistical performance to practical clinical implementation. A previous study investigating CR abnormality detection using AI reported that false-positive rates can significantly affect radiologist acceptance and workflow integration (28). Although the FRR observed in our study suggests acceptable clinical utility, institutional implementation requires careful threshold optimization tailored to clinical needs and tolerance for false alerts.

The retrospective, single-center design of our study also limits generalizability. Diagnostic performance may vary depending on institutional factors such as patient demographics, imaging protocols, and the prevalence of AAs. Furthermore, radiologist performance was assessed without full clinical context, which may have influenced comparative evaluations. Previous multi-reader studies have similarly emphasized the importance of clinical context in AI validation (29). In addition, the case-control design—focused on distinctly abnormal and normal CRs—may have overestimated AI performance compared with real-world populations that include borderline or equivocal findings. Prospective, multicenter validation is warranted to confirm external validity across diverse healthcare environments (30). In clinical practice, AI-based identification of AAs could function as a triage alert system, prioritizing critical CRs for prompt review. Safe implementation will require integration with radiology reporting systems and radiologist verification workflows (25,31).

In conclusion, the AI solution demonstrated high diagnostic performance in detecting AAs on CRs, achieving sensitivity and F1 scores comparable to or exceeding those of thoracic radiologists, albeit with a modestly higher FRR. These findings support the potential role of AI in assisting radiologists by enhancing detection consistency and workflow efficiency, provided that false alerts are effectively managed.

Supplementary Materials

The Supplement is available with this article at http://doi.org/10.3348/jksr.2025.0071.

Supplementary Table 1

Distribution of AI-Predicted Abnormalities and False-Positive Classifications