Artificial Intelligence‐Assisted Pulmonary Nodule Diagnosis by Thoracic Surgeons: A Comparative Study on Clinical Effectiveness Authors: Yeong Jeong Jeon, Jonghun Jeong, Doohyun Park, Jung‐Hyun Kang, Eun Kyoung Hong, Younjoon Chung, Boram Park, Seong Yong Park Published: 2026 ABSTRACT ABSTRACTMain ProblemClinician experience variability affects lung cancer detection using computed tomography. AI‐based computer‐aided diagnosis (CAD) systems can improve diagnostic accuracy, but their influence on thoracic surgeons with varying experience is unclear. This study aimed to evaluate the impact of CAD system on thoracic surgeons' assessment of pulmonary nodule malignancy.MethodsA multireader, crossover study included 20 thoracic surgeons (8 junior, 12 senior). Each reader interpreted 100 anonymized pulmonary nodules twice—once with CAD and once without—following a washout period. Nodules were assessed using a 10‐point likelihood of malignancy (LOM) scale and a binary (benign/malignant) assessment.ResultsThe CAD system demonstrated high malignancy prediction (AUROC = 0.929). CAD assistance significantly improved diagnostic accuracy across all readers (AUROC: 0.770–0.804, p = 0.0014 [BINARY]; 0.833–0.879, p < 0.001 [LOM]). Junior readers showed greater improvement (BINARY: 0.780–0.848, p < 0.001; LOM: 0.838–0.904, p < 0.001) compared with senior readers (BINARY: 0.763–0.775, p = 0.344; LOM: 0.831–0.862, p = 0.026). Specificity significantly increased with CAD (p = 0.008 overall, p < 0.001 junior readers), while sensitivity remained unchanged. Reading time slightly increased with CAD for benign cases (p < 0.05), but remained stable for malignant cases.ConclusionsThe CAD system improved thoracic surgeons' diagnostic performance, especially junior readers, by enhancing specificity and overall accuracy. AI‐based CAD systems are beneficial for reducing diagnostic variability and supporting lung nodule malignancy assessment by inexperienced thoracic surgeons. Introduction Lung cancer is a leading cause of cancer‐related mortality worldwide, mainly due to delayed diagnosis and difficulties in accurately identifying pulmonary nodules on chest computed tomography (CT) images. Early and precise detection of these nodules, especially smaller or subsolid lesions, is crucial for improving patient outcomes but continues to pose significant diagnostic challenges, even for experienced clinicians. Artificial intelligence (AI), particularly deep learning‐based computer‐aided diagnosis (CAD) systems, has demonstrated significant promise in assisting clinicians with lung nodule detection and characterization. Previous studies have convincingly demonstrated that AI can improve detection accuracy, enhance diagnostic confidence, and reduce observer variability among radiologists [1, 2, 3, 4, 5], pulmonologists [6], and advanced practice providers [7]. However, research investigating the utility and practical impact of AI tools among nonradiologist clinicians, particularly thoracic surgeons (who routinely interpret chest CT scans as part of their clinical practice), remains scarce [8, 9]. This gap is particularly relevant given that thoracic surgeons often review imaging independently during surgical evaluation, operative planning, and multidisciplinary conferences. Deep learning approaches for pulmonary nodule risk stratification have demonstrated promising results in several landmark studies. Massion et al. showed that deep learning models could accurately predict lung cancer risk from indeterminate pulmonary nodules in the National Lung Screening Trial (NLST), achieving performance comparable to or exceeding established clinical risk models [4]. Similarly, Mikhael et al. demonstrated that deep learning could predict future lung cancer risk from a single low‐dose CT scan [10]. These studies have established the technical feasibility of AI‐based malignancy prediction and provide important context for the present investigation. Despite recent technological advancements, several barriers hinder the widespread incorporation of AI into routine clinical workflows, including high false‐positive rates, challenges in integrating AI into established clinical routines, and uncertainty regarding the real‐world clinical utility of these systems [11, 12]. Moreover, performance variability across diverse clinical settings and patient populations underscores the necessity for further robust validation studies. While radiologists and pulmonologists serve as the primary interpreters in initial nodule detection and risk stratification, thoracic surgeons also regularly review chest CT imaging as part of comprehensive patient evaluation, surgical planning, and multidisciplinary tumor board discussions. In many institutions, thoracic surgeons independently assess imaging to determine surgical candidacy, plan operative approaches, and counsel patients. This is particularly relevant in settings where immediate radiology consultation may not be available, or when surgeons are evaluating imaging from outside institutions. Despite this regular imaging interpretation responsibility, research investigating AI utility specifically among thoracic surgeons has been scarce [8, 9], while most studies have focused on radiologists [1, 2, 3, 4, 5]. This study evaluated the diagnostic performance and clinical utility of AI‐assisted chest CT interpretation by thoracic surgeons. Specifically, it investigated the influence of AI assistance on diagnostic accuracy, clinical decision‐making efficiency, and diagnostic confidence among thoracic surgeons. By addressing this critical literature gap, this research offers valuable insights into AI's broader clinical applicability beyond radiology, aiming to inform future clinical practices in thoracic surgery and foster interdisciplinary collaboration for lung cancer management. Methods Study Design and Reader Cohort We recruited 20 thoracic surgeons (hereafter called readers) from two tertiary hospitals in South Korea. Based on years of clinical practice and involvement in chest imaging interpretation, these readers were categorized into: Junior readers (n = 8): ≤ 2 years of clinical training (considered less experienced). Specifically, this group included thoracic surgeons in their first or second year of thoracic surgery training and had limited specialized experience in thoracic imaging interpretation.Senior readers (n = 12): > 2 years of clinical training (considered more experienced). This group consisted of thoracic surgeons with more than 2 years of thoracic surgery training and were routinely involved in high‐volume thoracic oncology clinics and multidisciplinary tumor boards. Junior readers (n = 8): ≤ 2 years of clinical training (considered less experienced). Specifically, this group included thoracic surgeons in their first or second year of thoracic surgery training and had limited specialized experience in thoracic imaging interpretation. Junior readers (n = 8): ≤ 2 years of clinical training (considered less experienced). Specifically, this group included thoracic surgeons in their first or second year of thoracic surgery training and had limited specialized experience in thoracic imaging interpretation. Senior readers (n = 12): > 2 years of clinical training (considered more experienced). This group consisted of thoracic surgeons with more than 2 years of thoracic surgery training and were routinely involved in high‐volume thoracic oncology clinics and multidisciplinary tumor boards. Senior readers (n = 12): > 2 years of clinical training (considered more experienced). This group consisted of thoracic surgeons with more than 2 years of thoracic surgery training and were routinely involved in high‐volume thoracic oncology clinics and multidisciplinary tumor boards. To evaluate how CAD assistance influences diagnostic performance, we conducted a multi‐reader, crossover study with a 2‐month washout period to minimize recall bias. Each reader interpreted the same set of nodules twice: first in one condition (with or without CAD), then in the opposite condition after the washout. Datasets To ensure both the clinical relevance and reliability of nodule‐level annotations, we utilized the LIDC‐IDRI dataset, one of the most widely adopted public datasets in the field of lung nodule analysis. The LIDC‐IDRI dataset contains 1018 thoracic CT scans, each annotated by four experienced thoracic radiologists through a two‐phase review process. Nodules ≥ 3 mm in diameter were annotated and described further radiologic features such as spiculation, lobulation, and malignancy likelihood [13, 14]. For this study, we specifically selected the subset of 157 CT scans within LIDC‐IDRI that included metadata indicating the malignancy diagnosis at the patient level. These diagnosis method codes provide insight into how the original diagnosis was established (e.g., through radiologic follow‐up or pathology), and were used as a proxy for deriving ground truth labels in the absence of actual pathology reports [14]. Data were categorized into four diagnostic categories (1. Unknow, 2. Benign or non‐malignant disease, 3. Primary lung cancer, 4. Metastatic lesion associated with an extra‐thoracic primary malignancy) at the patient level based on the diagnostic method, which included the review of radiological images, biopsy, surgical resection, and assessment of disease progression or response. To define benign cases, we included cases diagnosed as “Benign or non‐malignant disease” based on one of the following diagnostic methods (1) radiological image review demonstrating a stable nodule over a period of 2 years, (2) biopsy, (3) surgical resection, or (4) assessment of disease progression or response. For malignant cases, we included those diagnosed as “Primary lung cancer” based on one of the following diagnostic methods (1) biopsy, (2) surgical resection, or (3) assessment of disease progression or response. To confirm the benign or malignant status at the nodule level, an in‐house radiologist with 10 years of experience in radiology reviewed all candidate chest CT cases. From this process, we identified 66 benign nodules from 31 CT scans, and 34 malignant nodules from 31 CT scans, yielding a balanced dataset of 100 nodules. This selection was subsequently reviewed by another in‐house radiologist with 14 years of experience in radiology to assess quality and potential confounders. As a result, five nodules were excluded for the following reasons: (1) severe calcification potentially leading to misinterpretation as benign, (2) the presence of a clearly malignant nodule appearing but unannotated nodule in the same axial slice that could bias reader judgment, or (3) with an unclear or insufficient diagnostic method, such as disease progression or response. To minimize the risk of misclassifying indolent malignancies as benign, all nodules deemed benign based on radiologic stability underwent independent review by two experienced radiologists (10 and 14 years of experience in thoracic imaging). Cases with any concerning features suggesting slow‐growing malignancy (e.g., subtle increase in attenuation, development of solid components, or irregular margins) were excluded from the benign category. Additionally, when extended follow‐up data beyond 2 years were available in the LIDC‐IDRI metadata, we verified continued stability. While we acknowledge that this approach cannot completely eliminate the possibility of misclassifying very indolent malignancies, these quality control measures substantially reduced this risk. Consequently, the final reader study dataset comprised 95 nodules, all manually verified to ensure diagnostic clarity and interpretive consistency. For each nodule, radiologic characteristics such as type (solid, part‐solid, or pure ground glass), location, and size were also recorded to support downstream subgroup analysis. AI Algorithm Development In the context of lung CT imaging, conventional computer‐aided detection (CADe) systems for nodule detection are typically trained using radiologist‐defined ground truth based on morphological characteristics (e.g., size, spiculation, nodule type) and guidelines such as Lung‐RADS [15]. In standard CADe workflows, labeling is often performed by multiple radiologists through independent review, followed by consensus on the classification of each nodule. While these ground truths reflect expert interpretation, they are inherently subjective [16, 17, 18]. In contrast, the CAD algorithm evaluated in this study was trained to directly estimate the likelihood of malignancy using biopsy‐ or surgery‐confirmed labels. Accordingly, it outputs a malignancy probability score. The reader study was conducted primarily with thoracic surgeons, rather than radiologists, to better align with the intended clinical application. This type of system is referred to as a CADx (Computer‐Aided Diagnosis), as it focuses on the characterization and diagnostic assessment of detected nodules [15]. For training, we utilized the National Lung Screening Trial (NLST) dataset [19], which contains cases with pathologically verified malignancy information. Following criteria consistent with the reader study, malignant nodules were selected based on biopsy or surgical resection confirmation, and benign nodules were selected based on documented radiologic stability. We acknowledge that benign labels in the NLST training dataset carry the same potential limitation, as they were similarly defined based on radiologic stability rather than pathological confirmation. However, this reflects a recognized limitation inherent to studies utilizing large‐scale public screening datasets and is consistent with the approach adopted by major prior deep learning studies in this field [4, 5, 10]. Given the natural imbalance in the dataset—with fewer pathologically confirmed malignancies—data augmentation techniques were applied to mitigate class imbalance during training. The final dataset consisted of 662 malignant nodules from 412 scans and 3622 benign nodules from 200 scans. A 5‐fold cross‐validation strategy was employed, and model outputs from each fold were combined through ensemble averaging to obtain final malignancy probability predictions. The model architecture comprises two stages. In Stage 1, a convolutional neural network (CNN) was used as a feature extractor, with EfficientNet V2 as the backbone. Feature maps were aggregated using Generalized Mean (GeM) pooling, and the loss function employed was weighted binary cross‐entropy (WBCE) to address class imbalance. In Stage 2, the slice‐level features extracted by the CNN were sequentially passed into a long short‐term memory (LSTM) module, which generated the final malignancy probability for each nodule. WBCE loss was also used in this stage to optimize learning. Key training details of the CADx model are summarized in Table S4. (Tables S1, S4, and Figure S1). Reader Study Protocol To evaluate the clinical impact of the CAD algorithm, we conducted a structured reader study using a cloud‐based deployment of the research‐use‐only version of VUNO Med—LungCA AI (hereafter referred to as LCA). This software system is designed to analyze chest CT scans, detect pulmonary nodules, and generate malignancy probability scores for each identified lesion. Although LCA is not a commercially approved product, it shares the same user interface as VUNO Med—LungCT AI, a CADe‐based software that has obtained regulatory approval in seven countries, including South Korea and Japan, and has been used in multiple published clinical studies. The LCA system was deployed on a dedicated Amazon Web Services (AWS) cloud environment for the purpose of this study. A total of 40 user accounts were created—two per reader (one with CAD and one without CAD). A dataset consisting of 58 CT scans with 95 selected nodules (as described in the Datasets section) was uploaded to the AWS instance. The LCA system successfully detected all 95 nodules and inferred corresponding malignancy probability scores. To minimize potential confusion during the study, we removed all auto‐detected nodules not included in the study dataset, such that only the results for the 95 ground truth nodules were visible to the readers. The 20 participating physicians were categorized as junior or senior readers based on whether they had more or less than 2 years of diagnostic training experience. To prevent inter‐reader influence, all patient information—including patient name and ID—was anonymized, and the case order was randomized differently for each account. Each reader was assigned both a with‐CAD and without‐CAD account. In the without‐CAD condition, the system displayed only detection information (e.g., nodule location, size), whereas in the with‐CAD condition, it additionally provided malignancy probability scores ranging from 0% to 100%. Readers were instructed to log in and complete the study independently via the web‐based platform. No discussion or result sharing among readers was allowed during the study period. The 20 readers were randomly divided into two crossover groups: Group 1 began the study using with‐CAD accounts in Stage 1, followed by a two‐month wash‐out period, and then completed Stage 2 using without‐CAD accounts.Group 2 followed the reverse order, starting with without‐CAD accounts in Stage 1 and switching to with‐CAD accounts in Stage 2 after the same wash‐out period. Group 1 began the study using with‐CAD accounts in Stage 1, followed by a two‐month wash‐out period, and then completed Stage 2 using without‐CAD accounts. Group 1 began the study using with‐CAD accounts in Stage 1, followed by a two‐month wash‐out period, and then completed Stage 2 using without‐CAD accounts. Group 2 followed the reverse order, starting with without‐CAD accounts in Stage 1 and switching to with‐CAD accounts in Stage 2 after the same wash‐out period. Group 2 followed the reverse order, starting with without‐CAD accounts in Stage 1 and switching to with‐CAD accounts in Stage 2 after the same wash‐out period. To further mitigate recall bias between stages, we not only applied the wash‐out period but also anonymized and shuffled the case order separately across each reader's two accounts, ensuring that case identifiers and sequences were different between conditions—even for the same reader. Figure 1 illustrates the overall workflow of the multi‐reader crossover study. Schematic of the workflow for the multi‐reader study. Each colored block represents one individual reader. Orange blocks indicate ‘with CAD assistance’ condition; blue blocks indicate ‘without CAD assistance’ condition. The entire reader study was conducted over approximately 6 months (April to September 2024), including the 2 months wash‐out periods. The reader study was completed, and data analysis began in October 2024. Statistical Analysis Each reader evaluated the reading results in two ways. The BINARY assessment categorized the findings as either malignancy or benign, while the likelihood of malignancy (LOM) was measured on a 10‐point scale (range: 1–10). A receiver operating characteristic (ROC) curve analysis was conducted to assess the diagnostic performance. The area under the ROC curve (AUROC) and 95% confidence interval (CI) was calculated based on both BINARY and LOM. A score of LOM ≥ 6 was considered indicative of suspected malignancy and was regarded as positive when calculating sensitivity and specificity. The difference in AUROC between without and with CAD was tested using Delong's method for two correlated ROC curves. The change in sensitivity and specificity was compared using McNemar's test. The difference of the reading times pooled across readers between without and with CAD assistance was tested using the Wilcoxon singed‐rank test and was presented using a box plot. A p value of less than 0.05 was considered statistically significant. All statistical analyses were performed with SAS 9.4 (SAS Institute, Cary, NC, USA) and R‐project for statistical computing (version 4.4.0). Results Characteristics Table 1 summarizes the final 95 nodules (65 benign, 30 malignant). Most benign nodules were classified as Lung‐RADS 1–3, while malignant nodules were notably present in categories 4B and 4X. Solid nodules were the most common type overall, but part‐solid nodules showed a higher proportion of malignancy (10/16 part‐solids). TABLE 1Nodule characteristics.CharacteristicTotalBenignMalignant(N = 95)(n = 65)(n = 30)Lung‐RADS144 (6.2%)0 (0.0%)23130 (46.2%)1 (3.3%)31717 (26.2%)0 (0.0%)4A75 (7.7%)2 (6.7%)4B152 (3.1%)13 (43.3%)4X217 (10.8%)14 (46.7%)Lobe locationLLL1714 (21.5%)3 (10.0%)LUL2919 (29.2%)10 (33.3%)RLL2015 (23.1%)5 (16.7%)RML88 (12.3%)0 (0.0%)RUL219 (13.8%)12 (40.0%)Nodule typeGGN76 (9.2%)1 (3.3%)Part‐solid166 (9.2%)10 (33.3%)Solid7253 (81.5%)19 (63.3%)Diameter (mm)Min–max3.6–32.33.6–27.59.6–32.3Mean ± SD11.6 ± 8.17.4 ± 4.720.7 ± 6.3Median (IQR)7.4 (5.3, 17.9)5.8 (4.8, 7.8)20.3 (16.4, 25.4)CAD‐generated malignancy probabilityMin–max0.3–99.90.3–99.847.8–99.9Mean ± SD (%)47.5 ± 37.829.2 ± 29.987.4 ± 16.6Median (IQR)47.8 (11.6, 91.1)14.6 (6.8, 50.2)95.5 (77.1, 99.1)Abbreviations: IQR, interquartile range; SD, standard deviation. Nodule characteristics. Abbreviations: IQR, interquartile range; SD, standard deviation. Mean nodule diameter was 11.6 ± 8.1 mm, with malignant nodules (20.7 ± 6.3 mm) significantly larger than benign nodules (7.4 ± 4.7 mm). The CAD system's predicted malignancy probability ranged from 0.3% to 99.9%, averaging 47.5% ± 37.8% across all nodules; malignant nodules had a higher mean score (87.4% ± 16.6%). Figure 2 presents four representative examples of lung cancer nodules for which CAD provided high malignancy probability scores. Examples of lung cancer detected with the aid of CAD. Diagnostic Performance of CAD‐Unassisted and CAD‐Assisted The CAD system measured the probability of malignancy, demonstrating a high predictive performance with an AUROC of 0.929 (95% CI, 0.880–0.978) (Figure 3). Internal validation using the NLST dataset yielded an AUROC of 0.840, indicating robust malignancy prediction performance. Diagnostic performance was evaluated for all readers (average across 20 readers), junior readers (average across 8 readers), and senior readers (average across 12 readers) based on the use of CAD in Table 2. When using BINARY for prediction, the use of CAD significantly improved predictive performance across all readers, increasing from 0.770 to 0.804 (p = 0.001). With CAD, the AUROCs changed from 0.780 to 0.848 in junior readers (p < 0.001) and from 0.763 to 0.775 in the senior readers (p = 0.344), with statistically significant differences, except for senior readers. When evaluating diagnostic performance using LOM, the use of CAD significantly increased predictive performance for all readers (0.833 vs. 0.879, p < 0.001), junior readers (0.838 vs. 0.904, p < 0.001), and senior readers (0.831 vs. 0.862, p = 0.026). The specificity also significantly increased for all readers (70.1% vs. 73.7%, p = 0.008) and junior readers (72.7% vs. 81.7%, p < 0.001). However, there was no significant difference in the sensitivity for all readers between CAD‐assisted and unassisted readings. Additionally, AUROC, sensitivity, and specificity were calculated for each individual reader (Tables S2 and S3). Performance of the CAD alone and junior and senior readers with and without CAD based on (A) BINARY and (B) likelihood of malignancy (LOM). TABLE 2Comparison of prediction performance between using and not using CAD.CharacteristicReader groupWithout CADWith CAD p AUROC (BINARY)All readers0.770 (0.750–0.789)0.804 (0.787–0.822)0.0014Junior readers0.780 (0.750–0.811)0.848 (0.822–0.875)0.0001Senior readers0.763 (0.738–0.787)0.775 (0.751–0.799)0.3438AUROC (LOM)All readers0.833 (0.813–0.852)0.879 (0.862–0.895)< 0.0001Junior readers0.838 (0.808–0.867)0.904 (0.881–0.928)0.0001Senior readers0.831 (0.805–0.857)0.862 (0.839–0.885)0.0264Sensitivity (%)All readers83.8 (80.9–86.8)87.2 (84.5–89.8)0.0521Junior readers83.3 (78.6–88.0)87.9 (83.8–92.0)0.1161Senior readers84.2 (80.4–87.9)86.7 (83.2–90.2)0.2332Specificity (%)All readers70.1 (67.6–72.6)73.7 (71.3–76.1)0.0075Junior readers72.7 (68.9–76.6)81.7 (78.4–85.0)< 0.0001Senior readers68.3 (65.1–71.6)68.4 (65.2–71.7)0.9438 Note: p‐values < 0.05 were considered statistically significant.Abbreviations: AUROC, area under the receiver operating characteristic curve; CAD, computer‐aided diagnosis; LOM, likelihood of malignancy. Comparison of prediction performance between using and not using CAD. Note: p‐values < 0.05 were considered statistically significant. Abbreviations: AUROC, area under the receiver operating characteristic curve; CAD, computer‐aided diagnosis; LOM, likelihood of malignancy. Effects of CAD Assistance on Reading Time The overall reading time pooled across readers increased for all nodules when assisted by CAD compared with reading without CAD (33 vs. 37.3 s, respectively). However, the reading time for malignant cases showed no statistically significant difference depending on CAD assistance (without CAD: 46.3 s vs. with CAD: 47.1 s). In comparison, the reading time for benign cases increased with CAD use (without CAD: 22.4 s vs. with CAD: 28.1 s). This trend was similarly observed in both junior and senior readers. There was no significant difference in reading time for malignant cases (40.8 s vs. 41.0 s for junior readers and 53.5 s vs. 53.6 s for senior readers). However, there was a significant difference in reading time for benign cases (18.3 s vs. 24.5 s for junior readers and 24.4 s vs. 29.6 s for senior readers) (Figure 4). Reading time pooled across readers without and with CAD assistance in (A) all readers, (B) junior readers, and (C) senior readers. Discussion This study provides the first comprehensive evaluation of AI‐assisted pulmonary nodule diagnosis specifically among thoracic surgeons, revealing significant improvements in diagnostic accuracy with the greatest benefits observed among less experienced surgeons. These findings have important implications for surgical education, clinical practice, and AI implementation in thoracic surgery. Although the majority of investigations in this field have focused on radiologists, our findings demonstrate that thoracic surgeons (and by extension, other nonradiologist clinicians) also derive meaningful benefits from AI‐driven CAD tools in the context of lung nodule evaluation. This aligns with earlier reports suggesting that individuals with comparatively limited imaging experience may benefit the most from AI support, particularly in demanding clinical environments where immediate radiology consultation may not always be feasible [20, 21]. Diagnostic Performance Improvements A key finding of this study is the increase in diagnostic specificity with the introduction of AI assistance. Specifically, the CAD system's capacity to highlight suspicious nodules appears to reduce the frequency of false‐positive classifications, which can otherwise lead to unnecessary follow‐up imaging, biopsies, or surgical interventions. AI tools may optimize patient management by assisting clinicians to more accurately rule out malignant nodules, reducing the psychological stress and healthcare costs associated with unnecessary concerns. These advantages align with prior studies reporting improvements in specificity when experienced radiologists used CAD [2, 22, 23, 24, 25, 26, 27], and our work suggests that similar benefits may extend to nonradiologist users. Interestingly, readers with different experience levels appeared to benefit from AI differently. Less experienced clinicians demonstrated more substantial accuracy improvements, likely because they had not yet reached the higher baseline performance typically observed among more experienced specialists. Thus, by offering consistent and objective guidance, AI systems helped less experienced clinicians identify subtle imaging markers that might otherwise have been missed. In contrast, more experienced readers showed only modest improvements, consistent with the ceiling effect observed in previous research [23]. Nevertheless, even modest specificity improvements are clinically valuable because they reduce unnecessary procedures. Impact on Clinical Workflow The incorporation of AI into clinical workflows poses new challenges. Specifically, our findings revealed a noticeable increase in the reading time for benign nodules when clinicians relied on AI‐generated malignancy probability scores. This is likely due to the additional verification steps clinicians feel compelled to undertake whenever the algorithm flags a lesion with a nonnegligible probability of malignancy. Although spending more time on truly suspicious nodules may seem logical, paying increased attention to those that are ultimately benign leads to a trade‐off because improved diagnostic accuracy increases the cognitive workload, thereby slowing the overall throughput. Similar phenomena were documented in radiology‐focused studies, where the time spent assessing inaccurate AI alerts or resolving unclear algorithmic findings occasionally exceeded the efficiency gains realized in more routine evaluations [3, 23, 25]. Clinical Implications for Thoracic Surgery The integration of AI‐assisted diagnosis into thoracic surgical practice offers several potential benefits. First, it may improve diagnostic consistency across surgeons with varying imaging experience, leading to more standardized patient care. Second, it could enhance confidence in surgical decision‐making, particularly for complex cases where malignancy probability is uncertain. Third, it may facilitate more informed discussions with patients and multidisciplinary teams regarding treatment recommendations. Our findings should be interpreted within the context of multidisciplinary lung cancer care. While radiologists and pulmonologists serve as primary interpreters in initial nodule detection and risk stratification, thoracic surgeons routinely encounter imaging interpretation responsibilities beyond the point of surgical referral—including review of outside institution studies, preoperative planning, intraoperative decision‐making, and multidisciplinary tumor board participation. In these contexts, immediate radiology consultation may not always be available, and surgeons must independently assess imaging to counsel patients and determine operative strategy. Our study specifically addresses this underexamined clinical scenario, rather than positioning AI as a replacement for established radiologic workflows. The demonstrated improvement in diagnostic accuracy—particularly among less experienced surgeons—suggests that AI assistance may help standardize nodule assessment across varying levels of surgical imaging expertise within the care continuum. However, successful implementation requires careful consideration of workflow integration, user training, and quality assurance measures. Surgeons must understand both the capabilities and limitations of AI tools to use them effectively in clinical practice. Our findings are consistent with the established literature demonstrating AI benefits across multiple clinician groups involved in lung nodule evaluation. Kim et al. [6] showed that CAD improved malignancy risk estimation and interobserver agreement among radiologists and pulmonologists, while Vachani et al. [7] demonstrated similar benefits for advanced practice providers. Our study extends these findings by demonstrating that thoracic surgeons—who typically evaluate patients later in the diagnostic pathway—also benefit meaningfully from AI assistance. This comprehensive evidence across multiple clinician groups suggests that AI‐assisted diagnosis may have broad applicability throughout the lung cancer care continuum, from initial detection through surgical evaluation. Limitations Despite these promising indications, certain limitations hinder the immediate, widespread adoption of AI tools. Subsolid and ground‐glass nodules remain a notable challenge for many current algorithms, resulting in higher false‐positive rates that erode user confidence and impede clinical workflow efficiency [11]. Several additional limitations must be considered when interpreting our findings. First, this study evaluated a single AI system (VUNO Med‐LungCA AI), and our results may not generalize to other commercially available or research‐based algorithms. Different AI systems employ varying architectures, training datasets, and optimization strategies, which can substantially affect diagnostic performance. Some systems focus on detection (CADe), while others emphasize characterization (CADx), and their clinical utility may differ accordingly. Clinicians and institutions considering AI implementation should evaluate multiple systems specific to their intended use case rather than assuming universal performance across all AI tools. Second, the nodule characteristics in our study dataset may not fully reflect real‐world clinical challenges. Specifically, malignant nodules in this study were significantly larger than benign nodules (20.7 ± 6.3 mm vs. 7.4 ± 4.7 mm), which may have facilitated detection and classification compared to smaller, more subtle lesions commonly encountered in lung cancer screening programs. This size differential could potentially overestimate both AI and physician performance when applied to populations with predominantly small nodules (< 8 mm), where diagnostic uncertainty is greatest. Future studies should specifically evaluate AI performance across different nodule size categories to better establish real‐world clinical utility. Third, a critical limitation of the current AI system is its inability to incorporate longitudinal imaging data. The CAD system assesses nodules at a single time point and cannot account for interval changes observed on serial imaging, such as growth kinetics, evolution of morphological features, or development of solid components within subsolid nodules. In clinical practice, temporal behavior is often the most important factor in determining malignancy risk, as even small nodules demonstrating growth over 3–6 month intervals warrant heightened suspicion, while stable nodules over extended periods are more likely benign. To our knowledge, no commercially available AI product currently integrates serial imaging data when generating malignancy probability scores, representing a significant gap between current AI capabilities and real‐world clinical decision‐making. Until such capabilities are developed, AI‐assisted nodule assessment should be viewed as complementary to, rather than replacing, longitudinal clinical evaluation. Future AI development should prioritize incorporating temporal analysis, enabling systems to quantify growth rates, track morphological evolution, and integrate these dynamic features into risk prediction. Fourth, although rigorously curated, the composition of the study population and the nature of the dataset may not completely mirror the heterogeneity of real‐world patient presentations and imaging protocols. Future Directions Future directions from these insights point to several critical research priorities. Wide‐ranging multi‐institutional studies encompassing different patient populations, imaging technologies, and clinical settings are necessary to confirm the broader applicability of AI‐based CAD tools. Additionally, algorithm enhancements specifically addressing subsolid and ground‐glass nodules will strengthen the reliability of automated characterization, particularly for early‐stage lung cancer screening initiatives. Interface design and educational approaches must also be refined to minimize the cognitive load on practitioners, ensuring that AI guidance enhances rather than disrupts established clinical processes. Finally, future studies are necessary to evaluate how AI‐assisted diagnosis ultimately affects patient results, resource efficiency, and cost to definitively establish the role of these technologies in comprehensive lung cancer management. Comparison With Established Risk Models It is essential to consider how AI‐driven CAD systems perform relative to established malignancy risk calculators, such as the Mayo Clinic model and the Brock University model [28, 29], which are widely utilized in clinical practice due to their accessibility and validated predictive power. Unlike these conventional models that primarily utilize patient demographics and nodule size, our CAD system leverages a deep‐learning architecture to analyze pixel‐level morphological characteristics, which may provide additional diagnostic insights, particularly for indeterminate nodules where clinical parameters alone are insufficient. However, future multi‐center studies should focus on a direct comparison between AI tools and traditional calculators to definitively establish the incremental clinical value and cost–benefit ratio of implementing proprietary AI platforms in thoracic surgical workflows. While our findings demonstrate the potential of AI‐assisted diagnosis, they are specific to the evaluated system, and performance characteristics of other AI platforms may differ considerably. Conclusions This research underscores how AI‐assisted CAD systems significantly enhance the diagnostic precision of thoracic surgeons and clinicians interpreting lung nodules amid practical limitations. The improved specificity these technologies provide, alongside their capacity to standardize assessments, suggests potential integration within comprehensive care frameworks. However, notable compromises, particularly longer reading times and potential operational disruptions, emphasize the need for ongoing technical refinements, intuitive design approaches, and pragmatic clinical testing. As AI becomes increasingly embedded within the medical infrastructure, the fundamental aim remains delivering more precise and individualized diagnostic processes that ultimately benefit patients with potential lung malignancies. While our findings demonstrate the potential of AI‐assisted diagnosis, they are specific to the evaluated system, and performance characteristics of other AI platforms may differ considerably. Author Contributions Jung‐Hyun Kang: validation, software, investigation. Doohyun Park: validation, software, investigation. Eun Kyoung Hong: software, validation, investigation. Yeong Jeong Jeon: writing – review and editing, writing – original draft, data curation. Younjoon Chung: software, validation, investigation. Seong Yong Park: conceptualization, writing – review and editing, project administration. Boram Park: methodology, formal analysis, writing – original draft. Jonghun Jeong: software, visualization, data curation. Funding This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (Ministry of Science and ICT, South Korea) (No. RS‐2025‐00553825) and by Vuno Inc., Republic of Korea. Ethics Statement This study used only publicly available, fully de‐identified datasets (NLST and LIDC‐IDRI). Accordingly, institutional review board approval and informed consent were not required for this study. Conflicts of Interest Jonghun Jeong, Doohyun Park, and Junghyun Kang are affiliated with Vuno Inc., Republic of Korea. The other authors declare no conflicts of interest. Supporting information Table S1: Model training scheme for malignancy‐classification CAD. Table S2: Comparison of AUROCC as prediction performance between without and with CAD. Table S3: Comparison of sensitivity and specificity as prediction performance between without and with CAD. Table S4: Key training details of the CADx model. Figure S1: AI algorithm model architecture.