Artificial intelligence in thoracic surgery: from nodule detection to surgical planning—a narrative review

Introduction

Background

Artificial intelligence (AI) is increasingly influencing clinical medicine by enabling advanced analysis of complex and high-dimensional data (1). In thoracic surgery, this evolution is occurring in parallel with major shifts in clinical practice, including the expansion of lung cancer screening (LCS) programs, the widespread adoption of minimally invasive and robotic techniques, and the growing reliance on imaging-based decision-making (2-5). These trends have substantially increased the volume and complexity of information that thoracic surgeons must integrate throughout the diagnostic, preoperative, and perioperative phases (Figure 1).

Figure 1 PubMed search using the terms “artificial intelligence” and “thoracic surgery”, showing over 4,000 results and an exponential increase in publications in recent years.

Recent advances in machine learning (ML) and deep learning (DL) have demonstrated promising results in pulmonary nodule detection, malignancy risk prediction, and radiological feature extraction (6). AI-driven tools are now being explored not only as diagnostic aids but also as decision-support systems for surgical planning, patient selection, and outcome prediction. As a result, AI is progressively moving from a purely research-oriented domain toward potential real-world clinical implementation within thoracic surgical workflows.

Rationale and knowledge gap

Despite the rapid growth of AI-related research, its translation into routine thoracic surgical practice remains limited. Current evidence is constrained by several limitations of primary studies, including retrospective design, heterogeneity of datasets, and limited external validation, which may affect generalizability across clinical settings (7-9). Existing studies often focus on isolated applications, such as nodule detection or radiomics-based prediction, without providing an integrated view of how AI can support the entire surgical pathway (10). Moreover, many reports emphasize technical performance metrics while offering limited discussion on clinical interpretability, workflow integration, and the evolving role of the thoracic surgeon as an active stakeholder in AI development and validation (10).

In addition, existing reviews often focus on isolated applications, such as nodule detection or radiomics-based prediction, without providing an integrated and clinically oriented perspective across the entire thoracic surgical pathway. Addressing this gap is essential to help surgeons critically appraise existing evidence, understand current limitations, and identify realistic opportunities for implementation in daily practice.

Objective

The objective of this narrative review is to provide a comprehensive and up-to-date overview of AI applications in thoracic surgery, with a particular focus on pulmonary nodule detection, surgical planning, and perioperative decision support. By summarizing current evidence and highlighting emerging trends, this review aims to offer thoracic surgeons a structured framework to understand the potential impact of AI on clinical decision-making and to identify key areas for future research and validation. This article is presented in accordance with the Narrative Review reporting checklist (available at https://asj.amegroups.com/article/view/10.21037/asj-2025-1-93/rc).

Methods

Literature search strategy

Given the broad and multidisciplinary scope of the topic, a narrative review approach was selected to allow integration of heterogeneous evidence across different domains of thoracic surgery. A non-systematic literature search was conducted using two electronic databases: PubMed/MEDLINE and Scopus. The search strategy combined Medical Subject Headings (MeSH) and free-text terms related to AI and thoracic surgical practice. Searches were limited to English-language publications.

The timeframe of interest ranged from January 2015 to December 2025, reflecting the period during which AI-based techniques began to show clinically meaningful applications in medical imaging and surgery.

Eligibility criteria

Eligible study types included original research articles, narrative and systematic reviews, technical reports, and expert opinions or commentaries with clinical relevance in the field. Exclusion criteria included editorials without substantive methodological or clinical discussion, conference abstracts, letters to the editor, and non-peer-reviewed material.

Study selection process

Study screening and selection were conducted by the author. Titles and abstracts were initially screened, followed by full-text assessment based on predefined inclusion and exclusion criteria, with a focus on clinical relevance to thoracic surgical practice and AI applications. This process was designed to ensure consistency in study selection while maintaining a clinically oriented approach. To improve transparency, the study selection process is summarized in Figure 2. No formal risk-of-bias assessment or quantitative synthesis was performed.

Figure 2 Flowchart illustrating the literature search and study selection process for this narrative review.

Data extraction and synthesis

Key information was extracted qualitatively from the included studies, focusing on the type of AI methodology used, the clinical context, and the potential impact on thoracic surgical practice. The literature was synthesized thematically and organized into conceptual domains. The methodological approach and search strategy are summarized in Table 1. A summary of the main studies included in this review, organized by application domain, is provided in Table 2. AI tools (ChatGPT, OpenAI) were used to support language editing and text organization during manuscript preparation.

AI in thoracic imaging and pulmonary nodule assessment

AI has emerged as a powerful tool in thoracic imaging, particularly in the context of pulmonary nodule detection and characterization. The increasing use of low-dose computed tomography (LDCT) in LCS programs has led to a substantial rise in imaging volumes, placing growing demands on radiologists and multidisciplinary teams. In this setting, AI-based systems have been proposed as decision-support tools to enhance detection accuracy, reduce workload, and standardize image interpretation across different clinical environments (9).

AI for pulmonary nodule detection in LDCT and clinical CT

Early computer-aided detection (CAD) systems have progressively evolved into DL-based architectures capable of automatically identifying pulmonary nodules on low-dose CT (LDCT) and standard chest CT scans. Recent systematic reviews consistently report high detection sensitivity, typically ranging between 86% and 98% for DL-based systems, with the highest performance observed for solid nodules and lesions ≥5 mm (11,25,26), albeit at the cost of lower specificity (77.5–87% vs. 87–91.7%) (11).

The ability of DL-based systems to process large image volumes rapidly is particularly relevant in LCS settings. However, most detection models have been trained and validated retrospectively, frequently using public datasets such as LIDC-IDRI and LUNA16. In a systematic review, 89% of included studies relied primarily on public datasets, and only 22% performed external validation, raising concerns about generalizability and real-world robustness (25).

Comparative studies between AI systems and radiologists suggest a complementary rather than competitive role. Another systematic review by Ewals et al. (12) showed that AI assistance increased radiologists’ sensitivity for nodule detection by 6.7 to 55.6 percentage points, while specificity was often modestly reduced. Importantly, when AI was used as a concurrent or second reader, radiologists’ area under the curve (AUC) improved by approximately 0.05–0.10, without a proportional increase in false positives in several studies. Similarly, Cheo et al. (11) reported that AI-assisted reading consistently improved sensitivity, while specificity loss varied widely depending on workflow integration.

From a screening and surgical perspective, high sensitivity must be carefully balanced against false-positive burden. False-positive rates for CT-based detection alone have been reported to reach 30–49%, even with AI assistance (11,27). Excessive detection of indeterminate nodules may therefore increase downstream follow-up imaging and invasive procedures. As highlighted by Kotoulas and colleagues, detection-only AI systems risk inappropriate surgical referral if not embedded within robust risk stratification frameworks (9). These findings underscore that AI-driven detection represents only the first step in the surgical pathway, and its clinical value depends critically on integration with malignancy prediction and decision-support tools rather than standalone deployment.

Malignancy risk prediction and radiomics

Radiomics represents a critical extension of AI applications beyond simple nodule detection, enabling the extraction of high-dimensional quantitative features related to shape, texture, intensity, and spatial heterogeneity. Multiple studies have demonstrated that CT-derived radiomic signatures can support malignancy risk prediction and, in selected settings, outperform conventional qualitative assessment by radiologists (13,26,28). Zhang et al. (14) reported that radiomics-based models achieved an AUC ranging from 0.78 to 0.86 for malignancy prediction, compared with 0.65–0.72 for visual assessment alone.

Evidence from systematic reviews and meta-analyses further supports the diagnostic potential of radiomics. A recent meta-analysis, including over 9,000 pulmonary nodules, reported pooled AUC values between 0.80 and 0.88 for ML- and DL-based radiomic models in differentiating benign from malignant nodules (13). However, diagnostic performance varied substantially across studies, with reported sensitivities ranging from 72% to 94% and specificities from 60% to 89%, reflecting heterogeneity in feature selection, modeling strategies, and reference standards (13,26).

Several investigations have explored the integration of radiomic features with clinical variables to enhance malignancy risk prediction. Hybrid models incorporating age, smoking history, and nodule size alongside radiomic features have consistently shown incremental predictive value. In particular, Zhang et al. (29) demonstrated that combined clinical-radiomic models improved AUC by approximately 0.05–0.10 compared with radiomics-only approaches. Similarly, another study reported that integrating radiomics with serum biomarkers and clinical features increased predictive accuracy, achieving AUC values up to 0.90 in internal validation cohorts (27).

Despite these advances, integration with widely used clinical risk models, such as PLCO-based approaches (30), remains inconsistent across studies. Most radiomics models are developed independently of established screening risk calculators and are rarely benchmarked against them, limiting their direct applicability in screening-derived surgical decision-making (25,26).

Reproducibility remains a major limitation. Variability in CT acquisition parameters, segmentation strategies, and feature extraction pipelines significantly affects model performance. In the systematic review by Palani et al. (26), fewer than 25% of studies performed external validation, and less than 20% provided sufficient methodological detail or code availability to allow reproducibility.

AI-assisted surgical planning in thoracic surgery

The progressive shift toward lung-sparing and minimally invasive procedures has increased the complexity of preoperative planning in thoracic surgery. In this context, AI-based tools have emerged as key enablers of precision surgery, supporting anatomical understanding, surgical strategy selection, and intraoperative navigation. Unlike earlier imaging-based applications, AI-assisted surgical planning aims to integrate patient-specific anatomy, tumor characteristics, and procedural constraints into a unified decision-support framework (9,31).

Preoperative planning and anatomical mapping

3D reconstruction represents one of the most mature and clinically adopted AI-assisted applications in thoracic surgical planning. By automatically segmenting pulmonary arteries, veins, bronchi, and tumor boundaries from thin-slice CT datasets, AI-driven platforms provide surgeons with a patient-specific, spatially accurate representation of bronchovascular anatomy. Evidence from meta-analyses indicates that AI-assisted 3D reconstruction improves anatomical recognition and intraoperative orientation compared with conventional bidimensional CT review, particularly in anatomically complex segmentectomies (15,16).

He and colleagues reported in a recent meta-analysis including over 1,200 patients that the use of 3D reconstruction was associated with a significant reduction in intraoperative complications and conversion rates, alongside improved identification of bronchovascular variants (15). Accuracy of AI-assisted anatomical matching between preoperative models and intraoperative findings has been reported to exceed 95–99% in multiple surgical series, even in complex segmentectomies (16,32). In the uniportal video-assisted thoracic surgery (VATS) setting, a 99.3% concordance rate between virtual 3D models and intraoperative bronchovascular anatomy was demonstrated, with anatomical variations identified in 36% of cases, underscoring the clinical relevance of preoperative mapping (32).

Virtual segmentectomy enabled by AI-based reconstruction allows simulation of resection planes, preoperative assessment of intersegmental margins, and anticipation of anatomical variants that may increase operative risk. These tools are particularly relevant in the context of sublobar resections, where margin adequacy and precise intersegmental plane identification are critical determinants of oncological efficacy (33). Several studies consistently report that access to 3D models improves surgeon confidence and reduces intraoperative uncertainty, especially in technically demanding procedures such as complex or combined segmentectomies (31,34).

Beyond visualization, AI-driven segmentation significantly reduces the time required for preoperative planning. Comparative studies indicate that AI-assisted reconstruction reduces model generation time from 20–30 minutes to less than 5 minutes, facilitating routine clinical adoption even in high-volume centers (16,35). Collectively, these findings support AI-assisted 3D reconstruction as a reliable and efficient tool for anatomical mapping, forming the foundation for more informed surgical strategy selection.

Surgical strategy selection

Beyond anatomical visualization, AI has been increasingly explored as a tool to support strategic surgical decisions, including the choice between lobectomy and segmentectomy. Several studies suggest that combining imaging-derived features with anatomical mapping may help identify patients suitable for sublobar resections while maintaining oncological adequacy (9,36). AI-assisted 3D reconstruction has been associated with changes in planned surgical strategy, including conversion from lobectomy to segmentectomy after detailed assessment of tumor location and margins (16). These findings highlight the potential role of AI in promoting lung preservation in appropriately selected patients.

Although prediction of nodal involvement remains primarily within the diagnostic domain, its implications for surgical planning are substantial. AI-based models integrating imaging features and anatomical data have shown promise in estimating resectability and guiding the extent of lymph node dissection (36). At present, however, AI should therefore be considered a complementary tool rather than a determinant of resectability, reinforcing and not replacing multidisciplinary evaluation.

Integration with minimally invasive and robotic surgery

The widespread adoption of VATS and robotic-assisted thoracic surgery (RATS) has increased procedural complexity and amplified the need for advanced preoperative planning and intraoperative guidance tools. In minimally invasive settings, where tactile feedback is reduced and the surgical field is limited, AI-assisted 3D reconstructions have been shown to improve surgeon orientation and anatomical awareness. In VATS segmentectomy, the use of AI-assisted 3D planning has been associated with reductions in operative time ranging from 10 to 25 minutes, as well as lower rates of unplanned conversion to thoracotomy or lobectomy (15,16).

In robotic surgery, AI-based segmentation and anatomical overlays can be integrated into console visualization systems, allowing surgeons to correlate real-time endoscopic views with preoperative anatomical maps. Early clinical experiences suggest that this integration may reduce intraoperative blood loss by improving vessel identification and dissection precision, with reported reductions of approximately 20–30% in selected series (17,36). Additionally, AI-assisted visualization has been associated with shorter docking and dissection times in complex robotic segmentectomies, although evidence remains limited to small, non-randomized cohorts (36).

Moreover, promising early feasibility results were reported, particularly in complex resections and deeply located nodules, with improved localization accuracy compared with conventional methods (16,34). In clinical practice, AI-assisted 3D reconstruction is increasingly integrated into preoperative planning workflows for VATS and robotic surgery, allowing surgeons to simulate resection strategies and anticipate anatomical variations before entering the operating room. These tools are already used in selected high-volume centers, particularly for complex segmentectomies, where precise anatomical understanding directly impacts surgical outcomes (37).

Despite rapid technological progress, several limitations currently hinder widespread adoption. Registration errors between virtual models and the intraoperative anatomy, lung deformation during ventilation, and workflow integration challenges remain significant barriers (9,31).

AI in perioperative risk stratification and outcome prediction

Perioperative risk stratification and outcome prediction represent a critical interface between surgical decision-making and patient-centered care in thoracic surgery. While traditional clinical scores remain widely adopted, they often fail to fully capture the complex, nonlinear interactions among patient characteristics, procedural factors, and perioperative physiology. AI-based models have therefore been proposed as tools to enhance individualized risk assessment and prognostication across the surgical pathway.

Prediction of postoperative morbidity and mortality

A growing body of evidence supports the application of ML models for predicting postoperative morbidity following thoracic surgery. These models typically integrate multidimensional perioperative data, including demographics, comorbidities, pulmonary function, laboratory parameters, and intraoperative variables, achieving superior discriminative performance compared with traditional statistical approaches (9,38). From a practical standpoint, ML-based risk prediction models may be integrated into electronic medical record (EMR) systems to provide real-time perioperative risk stratification. Such tools could support preoperative counseling, patient selection, and postoperative management by identifying high-risk individuals who may benefit from tailored perioperative strategies (31).

Several studies have specifically focused on the prediction of postoperative cardiopulmonary complications. In a retrospective cohort of lung resection patients, ML-based models achieved an AUC of approximately 0.77, outperforming conventional logistic regression approaches (39). Similarly, Zhai et al. (19) demonstrated improved predictive accuracy for cardiopulmonary morbidity using ML algorithms, with reported AUC values in the 0.75–0.80 range, highlighting the ability of these models to capture nonlinear interactions among clinical variables. More recent investigations have extended these findings to minimally invasive surgery. In a large thoracoscopic cohort exceeding 1,600 patients, an explainable ML model for cardiopulmonary complication prediction was developed, achieving an AUC of 0.82 in internal validation, 0.73 in temporal validation, and 0.76 in external validation (40). Importantly, this study incorporated intraoperative variables and demonstrated that ML-based risk stratification remained robust across different validation settings, supporting its potential applicability in real-world VATS populations. Comparable performance has been reported in other ML-based postoperative pulmonary complication prediction studies focusing on thoracoscopic procedures (19,39,41).

When directly compared with traditional clinical risk scores, AI-based models consistently demonstrate superior predictive performance. Mahajan et al. (18) showed that ML algorithms achieved higher AUC values than conventional clinical indices for adverse event prediction, particularly in heterogeneous surgical cohorts. Similarly, in patients undergoing major lung resections, a single-center retrospective study reported that ML-based models improved discrimination for postoperative complications compared with established clinical scores, with AUC improvements of approximately 0.05–0.10 over traditional approaches (42). These performance gains are largely attributable to the ability of ML models to account for complex, nonlinear relationships and high-order interactions that are not captured by conventional scoring systems (9,38).

Long-term outcomes and oncologic prognostication

Beyond short-term perioperative outcomes, AI has been increasingly applied to predict long-term oncologic endpoints following lung cancer surgery. ML and radiomics-based models have demonstrated promising performance in predicting disease-free survival (DFS), overall survival (OS), and recurrence risk, frequently outperforming TNM staging alone.

In a retrospective cohort study, ML models incorporating CT-derived imaging features achieved an AUC of approximately 0.80 for recurrence prediction, compared with 0.65–0.70 using TNM staging alone, highlighting the added prognostic value of imaging-based AI (21). Similar findings were reported by Kinoshita et al. (20), who developed ML-based prognostic models integrating radiomic and clinical variables to stratify long-term survival after curative lung resection. In their study, AI-derived risk stratification demonstrated a C-index of 0.78–0.82 for OS prediction, significantly outperforming conventional clinicopathological models. These results support the ability of AI to capture complex interactions between tumor phenotype and host factors that are not reflected in anatomical staging.

More investigations have focused on hybrid approaches combining radiomic features with clinical and pathological variables. A recent study demonstrated that integrating peritumoral radiomics with clinical data improved recurrence prediction, achieving AUC values up to 0.84, particularly in early-stage disease (43). Similarly, Lan et al. (22) reported that a 2.5D peritumoral radiomics model achieved an AUC of 0.83 for recurrence prediction in stage I lung adenocarcinoma, outperforming models based on intratumoral features alone. These results were further confirmed in another study showing that comprehensive radiomics-clinical models provide robust recurrence prediction across different histological subtypes, with reported C-index values exceeding 0.80 in internal validation cohorts (44).

Several contemporary studies have refined recurrence prediction by integrating both preoperative and postoperative data. Mehri-Kakavand and colleagues reported that ML-based models combining radiomic features with pathological variables improved recurrence risk stratification beyond pathological staging alone, with AUC improvements of approximately 0.08–0.12 compared with TNM-based models (45). Similarly, a retrospective European study demonstrated that AI-based models outperformed both clinical and pathological staging in predicting recurrence in pre- and postoperative settings, particularly when longitudinal data were incorporated (46).

Despite these encouraging results, it has to be considered that fewer than 30% of prognostic AI studies include external validation, and standardized feature extraction and reporting remain uncommon (9,38). Postoperative pathological factors (such as lymph node status) remain the strongest predictors of outcome, underscoring that AI-based prognostication is most effective when integrated with established oncologic parameters rather than used in isolation.

Integration of multimodal data: toward a surgical AI ecosystem

Imaging, clinical data, and biomarkers

The integration of heterogeneous data sources represents a critical step toward clinically meaningful AI applications in thoracic surgery. While most current AI models rely on single-modality inputs (predominantly CT imaging), surgical decision-making in lung cancer inherently requires the synthesis of imaging findings, clinical variables, pathological features, and molecular biomarkers (7,47). Over 70% of the published AI studies in lung cancer remain unimodal, limiting their ability to reflect real-world clinical complexity (7,24).

Radiomics has emerged as a foundational component of multimodal integration by enabling high-dimensional quantitative characterization of tumors beyond conventional semantic descriptors. Radiomic features correlate with histological subtype, tumor grade, molecular alterations (including EGFR and KRAS status), and survival outcomes (48-50). Reported performance metrics for radiomics-based prognostic models frequently reach C-index or AUC values in the range of 0.75–0.85, exceeding those of clinical models alone in retrospective cohorts (48,50).

Radiogenomics further extends this paradigm by linking imaging phenotypes with genomic and transcriptomic profiles, offering a non-invasive surrogate of tumor biology. Several CT-derived radiomic patterns and gene-expression signatures were found to be related to proliferation, hypoxia, and immune infiltration, with potential implications for surgical stratification and patient selection (24,47,49).

Beyond CT-based radiomics, multimodal imaging, and particularly PET-derived radiomics, adds functional and metabolic information that enhances prognostic modeling. Combining PET and CT radiomic features improves risk stratification and outcome prediction compared with either modality alone, with incremental AUC gains of approximately 0.05–0.10 in prognostic models (50,51). Heterogeneity in acquisition protocols, segmentation strategies, and feature extraction pipelines continues however to limit reproducibility and cross-center generalizability (24,49).

Liquid biopsy represents a complementary and dynamic data layer that captures tumor heterogeneity and temporal evolution. Circulating tumor DNA (ctDNA) has gained increasing relevance across screening, diagnosis, minimal residual disease detection, and postoperative surveillance. Thalambedu et al. (23) reported that AI-enhanced ctDNA analysis enables detection of mutant allele fractions as low as 0.002%, achieves relapse detection sensitivities up to 94%, and identifies recurrence with a median lead time of approximately 5.2 months before radiological evidence. These findings may inform perioperative risk stratification, early detection of recurrence, and longitudinal follow-up after curative-intent surgery, overcoming the limitations of individual modalities, particularly in early-stage disease where sensitivity remains suboptimal (23,52).

From isolated tools to end-to-end surgical pathways

Despite rapid advances in AI methodologies, most applications in thoracic oncology remain fragmented and confined to isolated steps of the clinical workflow, such as nodule detection, malignancy prediction, or postoperative risk assessment (7,8). This compartmentalized development contrasts with the continuous and iterative nature of the thoracic surgical pathway, which spans screening, diagnosis, surgical planning, intraoperative decision-making, and long-term surveillance (9).

AI models are frequently developed in silos, using narrowly defined datasets and single endpoints, which limits interoperability across clinical domains (8,31). As a result, AI tools are rarely embedded into real-world surgical workflows, and their clinical impact remains modest despite strong performance metrics in retrospective or controlled settings. Xu et al. (24) noted that fewer than 25% of published multimodal AI studies explicitly address clinical integration or workflow deployment.

Conceptual frameworks for an integrated surgical AI ecosystem advocate a shift toward end-to-end architectures capable of aggregating multimodal data, including imaging, electronic health records, pathological findings, and molecular biomarkers such as ctDNA, into unified, adaptive decision-support systems (23,24). In this model, AI would function not as a standalone diagnostic tool but as a longitudinal companion to the surgical pathway, continuously updated through perioperative data and postoperative outcomes (9).

Addressing the discussed limitations in this field will be essential to transition from proof-of-concept studies toward clinically deployable AI platforms that meaningfully support thoracic surgeons across the entire continuum of care (9,31).

Ethical, legal, and educational challenges

The integration of AI into thoracic surgery raises substantial ethical, legal, and educational challenges that must be addressed to ensure safe, transparent, and responsible clinical adoption. A central concern is explainability. Many high-performing AI systems, particularly DL models used for imaging analysis, risk prediction, and decision support, operate as “black boxes”, offering limited insight into how outputs are generated. Several authors consistently identify lack of interpretability as one of the main barriers to clinical trust and adoption, especially in high-stakes surgical settings where accountability and shared decision-making are essential (7,9,53). In clinical practice, limited explainability of AI outputs may affect multidisciplinary tumor board discussions, where treatment decisions require transparent justification (53,54). For example, an AI model suggesting high malignancy risk without clearly interpretable features may reduce clinician confidence and influence the decision to proceed with invasive diagnostic or surgical interventions. Similarly, algorithmic bias may impact risk estimation across different patient subgroups, potentially affecting surgical candidacy and informed consent (7,8). Explainable AI techniques, such as feature attribution and saliency mapping, have been proposed to clarify the relative contribution of clinical variables or imaging features to algorithmic predictions, enabling surgeons to critically appraise AI outputs rather than relying on them uncritically (36,54).

Algorithmic bias and data governance represent additional major concerns. Many AI models in thoracic oncology are trained on retrospective, single-center, or demographically unbalanced datasets, increasing the risk of performance disparities across sex, age, ethnicity, and comorbidity profiles (7,8). These limitations raise the possibility of reinforcing existing healthcare inequalities rather than mitigating them. At the same time, the demand for large, multimodal datasets conflicts with increasingly stringent data-protection regulations governing patient privacy, data sharing, and secondary use of clinical information. Ethical analyses emphasize that responsible AI deployment requires robust governance frameworks, transparent validation processes, and continuous post-deployment monitoring to ensure fairness, safety, and equity (9). Legal accountability remains incompletely defined, particularly when AI-generated recommendations influence surgical planning or intraoperative decisions, reinforcing the necessity of maintaining a clearly defined human-in-the-loop paradigm.

The expanding role of AI in thoracic surgery also has significant implications for surgical education and training. Satapathy et al. (55) consistently stressed that future thoracic surgeons will require not only technical proficiency in minimally invasive and robotic techniques but also foundational literacy in AI principles, including model limitations, uncertainty, and bias. Educational applications of AI such as simulation, performance assessment, and personalized learning, offer clear opportunities but also raise ethical concerns related to overreliance on automated feedback and erosion of independent clinical judgment (55,56). Without structured educational frameworks that integrate AI literacy, ethical reasoning, and critical appraisal skills, there is a risk that AI tools will be either uncritically accepted or inadequately utilized, ultimately limiting their potential benefit for patient care (9,56).

From a clinical perspective, ensuring the reliability of AI systems represents a critical prerequisite for their safe implementation in thoracic surgery. External validation on independent and heterogeneous cohorts remains essential to assess robustness and reproducibility before clinical deployment. Data quality and standardization also play a central role, as variability in imaging protocols, annotation strategies, and clinical data collection can significantly affect model performance (9). In addition, AI systems are subject to performance degradation over time due to changes in clinical practice, patient characteristics, or imaging technologies, a phenomenon often referred to as model drift. Continuous performance monitoring and periodic recalibration are therefore necessary to maintain reliability in real-world settings (9,57). Finally, the integration of AI into clinical workflows should be framed within a human-in-the-loop model, in which algorithmic outputs complement rather than replace clinical judgment. This approach ensures that AI-driven recommendations are interpreted in the context of patient-specific factors, thereby reducing the risk of inappropriate decisions and enhancing overall safety (9,36).

Future perspectives

The future impact of AI in thoracic surgery will depend less on incremental improvements in algorithmic performance and more on rigorous clinical validation, meaningful integration into surgical workflows, and active involvement of surgeons in AI development. Despite the growing number of retrospective studies and proof-of-concept models, prospective validation remains limited. Few AI tools have been tested within randomized or pragmatic clinical trials capable of demonstrating real-world benefit in terms of patient outcomes, workflow efficiency, or decision quality (7,9,31). Future research should therefore prioritize prospective, multicenter studies with standardized endpoints and transparent reporting to move beyond technical feasibility toward clinical utility.

A second critical direction involves surgeon-driven AI development. Many current AI applications are designed primarily by data scientists and engineers, often without sufficient alignment with surgical decision-making processes. Meaningful progress requires surgeons to act not only as end users but also as co-developers and clinical validators, ensuring that AI tools address relevant clinical questions, incorporate domain knowledge, and fit within real operative constraints (8,36). Such collaboration is essential to avoid the proliferation of isolated tools that perform well in silico but fail to influence surgical practice. Importantly, AI should be framed as a decision-support technology rather than a replacement for surgical judgment. Across imaging, surgical planning, perioperative risk stratification, and prognostication, the most consistent benefit of AI emerges when algorithms augment human expertise by synthesizing complex data and highlighting patterns that may not be readily apparent, while leaving final decisions to clinicians (9,53,57). Overreliance on automated outputs risks eroding professional autonomy and accountability, whereas a human-in-the-loop paradigm preserves ethical integrity and patient-centered care.

Strengths and limitations

This narrative review provides a comprehensive and clinically oriented synthesis of AI applications across the thoracic surgical pathway, moving beyond isolated technical solutions to offer an integrated, surgeon-centered perspective. By addressing imaging, surgical planning, perioperative risk stratification, multimodal data integration, and ethical considerations within a single framework, it reflects the real-world complexity of thoracic surgical decision-making. Moreover, the focus on clinically meaningful endpoints—such as surgical strategy selection, complication risk assessment, and oncologic outcomes—enhances the translational relevance of the discussed evidence. Finally, the structured narrative approach allows critical appraisal of current achievements and limitations, highlighting unmet needs and future research priorities from the standpoint of practicing thoracic surgeons.

Several limitations should also be acknowledged. As a narrative review, this work does not aim to provide an exhaustive or quantitative synthesis of the available literature, and no formal risk-of-bias assessment or meta-analysis was performed. The selection of studies may therefore be subject to selection bias. In addition, most of the available evidence in this field is based on retrospective studies with limited external validation, often derived from single-center datasets. This may affect the generalizability and robustness of reported results across different clinical settings. These limitations should be considered when interpreting the findings of this review. Finally, the rapidly evolving nature of AI research means that some technologies discussed here may undergo significant refinement or validation in the very near future.

Conclusions

AI is increasingly shaping the landscape of thoracic surgery, offering new opportunities to support clinical decision-making across the entire surgical pathway. The evidence summarized in this narrative review highlights that AI-driven tools can enhance the analysis of complex and high-dimensional data, improve anatomical understanding, and support more personalized surgical strategies. However, their current clinical impact remains uneven, with most applications still confined to isolated tasks rather than fully integrated workflows.

The greatest value of AI emerges when algorithmic outputs are interpreted within a multidisciplinary and clinically informed context, complementing experience, anatomical knowledge, and patient-centered reasoning. Surgeon engagement in the development, validation, and implementation of AI tools is therefore essential to ensure that these technologies address relevant clinical questions and integrate seamlessly into real-world practice. If guided by clinical needs, ethical governance, and interdisciplinary collaboration, AI has the potential to evolve from a collection of promising tools into a cohesive surgical ecosystem that supports safer, more precise, and more individualized thoracic surgical care.

Acknowledgments

Artificial intelligence tools (ChatGPT, OpenAI) were used to support language editing and text organization during manuscript preparation. No AI tools were used for data collection, analysis, or figure generation. The author is fully responsible for the content of the manuscript.

Footnote

Reporting Checklist: The author has completed the Narrative Review reporting checklist. Available at https://asj.amegroups.com/article/view/10.21037/asj-2025-1-93/rc

Peer Review File: Available at https://asj.amegroups.com/article/view/10.21037/asj-2025-1-93/prf

Funding: None.

Conflicts of Interest: The author has completed the ICMJE uniform disclosure form (available at https://asj.amegroups.com/article/view/10.21037/asj-2025-1-93/coif). The author has no conflicts of interest to declare.

Ethical Statement: The author is accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Obermeyer Z. Bedside to Bench - AI and the New Science of Medicine. N Engl J Med 2025;393:2287-89. [Crossref] [PubMed]
2. Lung Cancer Policy Network. Interactive map of lung cancer screening. Accessed December 21, 2025. Available online: https://www.lungcancerpolicynetwork.com/interactive-map-of-lung-cancer-screening/
3. Muriana P, Rossetti F, Novellis P, et al. Lung Cancer Screening: The European Perspective. Thorac Surg Clin 2023;33:375-83. [Crossref] [PubMed]
4. American College of Radiology Committee on Lung-RADS®. Lung-RADS Assessment Categories 2022. Accessed September 29, 2023. Available online: https://www.acr.org/-/media/ACR/Files/RADS/Lung-RADS/LungRADSAssessmentCategoriesv1-1.pdf
5. Potter AL, Spasojevic A, Raman V, et al. The Increasing Adoption of Minimally Invasive Lobectomy in the United States. Ann Thorac Surg 2023;116:222-9. [Crossref] [PubMed]
6. Yuan X, Xu H, Zhu J, et al. Systematic review and meta-analysis of artificial intelligence for image-based lung cancer classification and prognostic evaluation. NPJ Precis Oncol 2025;9:300. [Crossref] [PubMed]
7. Ladbury C, Amini A, Govindarajan A, et al. Integration of artificial intelligence in lung cancer: Rise of the machine. Cell Rep Med 2023;4:100933. [Crossref] [PubMed]
8. Liao W, Peng B, Ding G, et al. Artificial intelligence in thoracic surgery: Perspectives and challenges. Surg Pract Sci 2025;23:100317. [Crossref] [PubMed]
9. Kotoulas SC, Spyratos D, Porpodis K, et al. A Thorough Review of the Clinical Applications of Artificial Intelligence in Lung Cancer. Cancers (Basel) 2025;17:882. [Crossref] [PubMed]
10. Gritsiuta AI, Su S, Petrov RV. Navigating the challenges of artificial intelligence integration in thoracic surgery. Curr Chall Thorac Surg 2025;7:31. [Crossref] [PubMed]
11. Cheo HM, Ong CYG, Ting Y. A Systematic Review of AI Performance in Lung Cancer Detection on CT Thorax. Healthcare (Basel) 2025;13:1510. [Crossref] [PubMed]
12. Ewals LJS, van der Wulp K, van den Borne BEEM, et al. The Effects of Artificial Intelligence Assistance on the Radiologists’ Assessment of Lung Nodules on CT Scans: A Systematic Review. J Clin Med 2023;12:3536. [Crossref] [PubMed]
13. Wulaningsih W, Villamaria C, Akram A, et al. Deep Learning Models for Predicting Malignancy Risk in CT-Detected Pulmonary Nodules: A Systematic Review and Meta-analysis. Lung 2024;202:625-36. [Crossref] [PubMed]
14. Zhang R, Wei Y, Shi F, et al. The diagnostic and prognostic value of radiomics and deep learning technologies for patients with solid pulmonary nodules in chest CT images. BMC Cancer 2022;22:1118. [Crossref] [PubMed]
15. He H, Yu C, Yang Y, et al. Three-Dimensional Reconstruction and Virtual Simulation of Patient-Specific Anatomy for Procedural Planning in Thoracoscopic Segmentectomy: A Systematic Review and Meta-Analysis. Eur J Cardiothorac Surg 2025;67:ezaf283. [Crossref] [PubMed]
16. Song Z, Izhar A, Wei B. Artificial intelligence-assisted three-dimensional reconstruction in thoracic surgery: a narrative review. Curr Chall Thorac Surg 2025;7:6.
17. Sadeghi AH, Mank Q, Tuzcu AS, et al. Artificial intelligence-assisted augmented reality robotic lung surgery: Navigating the future of thoracic surgery. JTCVS Tech 2024;26:121-5. [Crossref] [PubMed]
18. Mahajan A, Esper S, Oo TH, et al. Development and Validation of a Machine Learning Model to Identify Patients Before Surgery at High Risk for Postoperative Adverse Events. JAMA Netw Open 2023;6:e2322285. [Crossref] [PubMed]
19. Zhai Y, Lin X, Wei Q, et al. Interpretable prediction of cardiopulmonary complications after non-small cell lung cancer surgery based on machine learning and SHapley additive exPlanations. Heliyon 2023;9:e17772. [Crossref] [PubMed]
20. Kinoshita F, Takenaka T, Yamashita T, et al. Development of artificial intelligence prognostic model for surgically resected non-small cell lung cancer. Sci Rep 2023;13:15683. [Crossref] [PubMed]
21. Shimada Y, Kudo Y, Maehara S, et al. Radiomics with Artificial Intelligence for the Prediction of Early Recurrence in Patients with Clinical Stage IA Lung Cancer. Ann Surg Oncol 2022;29:8185-93. [Crossref] [PubMed]
22. Lan H, Wei C, Xu F, et al. 2.5D peritumoural radiomics predicts postoperative recurrence in stage I lung adenocarcinoma. Front Oncol 2024;14:1382815. [Crossref] [PubMed]
23. Thalambedu N, Balla M, Sivasubramanian BP, et al. Integrating artificial intelligence with circulating tumor DNA for non-small cell lung cancer: opportunities, challenges, and future directions. Front Med (Lausanne) 2025;12:1612376. [Crossref] [PubMed]
24. Xu X, Li J, Zhu Z, et al. A Comprehensive Review on Synergy of Multi-Modal Data and AI Technologies in Medical Diagnosis. Bioengineering (Basel) 2024;11:219. [Crossref] [PubMed]
25. Gao C, Wu L, Wu W, et al. Deep learning in pulmonary nodule detection and segmentation: a systematic review. Eur Radiol 2025;35:255-66. [Crossref] [PubMed]
26. Palani M, Rajagopal S, Chintanpalli AK. A systematic review on feature extraction methods and deep learning models for detection of cancerous lung nodules at an early stage -the recent trends and challenges. Biomed Phys Eng Express 2024;
27. Ding Y, Zhang J, Zhuang W, et al. Improving the efficiency of identifying malignant pulmonary nodules before surgery via a combination of artificial intelligence CT image recognition and serum autoantibodies. Eur Radiol 2023;33:3092-102. [Crossref] [PubMed]
28. Zhang Y, Jiang B, Zhang L, et al. Lung Nodule Detectability of Artificial Intelligence-assisted CT Image Reading in Lung Cancer Screening. Curr Med Imaging 2022;18:327-34. [Crossref] [PubMed]
29. Zhang H, Wang S, Deng Z, et al. Computed tomography-based radiomics machine learning models for prediction of histological invasiveness with sub-centimeter subsolid pulmonary nodules: a retrospective study. PeerJ 2023;11:e14559. [Crossref] [PubMed]
30. Tammemägi MC, Katki HA, Hocking WG, et al. Selection criteria for lung-cancer screening. N Engl J Med 2013;368:728-36. [Crossref] [PubMed]
31. Leivaditis V, Maniatopoulos AA, Lausberg H, et al. Artificial Intelligence in Thoracic Surgery: A Review Bridging Innovation and Clinical Practice for the Next Generation of Surgical Care. J Clin Med 2025;14:2729. [Crossref] [PubMed]
32. Gürz S, Şengül A. Three-dimensional modeling of bronchovascular anatomy for preoperative planning and intraoperative guidance in uniportal video-assisted thoracoscopic segmentectomy. Turk Gogus Kalp Damar Cerrahisi Derg 2025;33:341-8. [Crossref] [PubMed]
33. Nakagawa K, Watanabe SI, Wakabayashi M, et al. Risk Factors for Locoregional Relapse After Segmentectomy: Supplementary Analysis of the JCOG0802/WJOG4607L Trial. J Thorac Oncol 2025;20:157-66. [Crossref] [PubMed]
34. Pohlman A, Hallare J, Facktor MA, et al. 3-dimensional reconstruction and mixed reality in thoracic surgery: a narrative review and user guide. J Thorac Dis 2025;17:11402-19. [Crossref] [PubMed]
35. Mank QJ, Thabit A, Maat APWM, et al. State-of-the-art artificial intelligence methods for pre-operative planning of cardiothoracic surgery and interventions: a narrative review. J Thorac Dis 2025;17:5282-97. [Crossref] [PubMed]
36. Rahouma M, Mohsen H, Mahmoud A, et al. Artificial intelligence (AI) applications and their impact on thoracic surgery: a narrative review. Curr Chall Thorac Surg 2025;7:27.
37. Lee CY, Yu WS, Yang YH, et al. Patients-specific virtual surgical navigation for lung segmentectomy: a prospective multicenter study. Front Oncol 2026;16:1774607. [Crossref] [PubMed]
38. Li J, Wu J, Zhao Z, et al. Artificial intelligence-assisted decision making for prognosis and drug efficacy prediction in lung cancer patients: a narrative review. J Thorac Dis 2021;13:7021-33. [Crossref] [PubMed]
39. Huang G, Liu L, Wang L, et al. Prediction of postoperative cardiopulmonary complications after lung resection in a Chinese population: A machine learning-based study. Front Oncol 2022;12:1003722. [Crossref] [PubMed]
40. Wang S, Lin Y, Shi H, et al. Prediction model for postoperative pulmonary complications after thoracoscopic surgery with machine learning algorithms and SHapley Additive exPlanations. J Thorac Dis 2025;17:3603-18. [Crossref] [PubMed]
41. Chen S, Deng T, Yang Q, et al. Development and validation of an explainable machine learning model for predicting postoperative pulmonary complications after lung cancer surgery: a machine learning study. EClinicalMedicine 2025;86:103386. [Crossref] [PubMed]
42. Hernandez MC, Chen C, Nguyen A, et al. Explainable Machine Learning Model to Preoperatively Predict Postoperative Complications in Inpatients With Cancer Undergoing Major Operations. JCO Clin Cancer Inform 2024;8:e2300247. [Crossref] [PubMed]
43. Ho WLJ, Fetisov N, Hall LO, et al. Evaluating clinical and radiomic features for predicting lung cancer recurrence pre- and post-tumor resection. Proc SPIE Int Soc Opt Eng 2024;12926:1292623. [Crossref] [PubMed]
44. Wang P, Luo Z, Luo C, et al. Application of a Comprehensive Model Based on CT Radiomics and Clinical Features for Postoperative Recurrence Risk Prediction in Non-small Cell Lung Cancer. Acad Radiol 2024;31:2579-90. [Crossref] [PubMed]
45. Mehri-Kakavand G, Mdletshe S, Amini M, et al. Multimodal radiomics fusion for predicting postoperative recurrence in NSCLC patients. J Cancer Res Clin Oncol 2025;151:261. [Crossref] [PubMed]
46. Valter A, Kordemets T, Gasimova A, et al. Pre- and post-operative lung cancer recurrence prediction following curative surgery: A retrospective study using European radiomics and clinical data. J Clin Oncol 2024;42:8066.
47. Wong CW, Chaudhry A. Radiogenomics of lung cancer. J Thorac Dis 2020;12:5104-9. [Crossref] [PubMed]
48. Tunali I, Gillies RJ, Schabath MB. Application of Radiomics and Artificial Intelligence for Lung Cancer Precision Medicine. Cold Spring Harb Perspect Med 2021;11:a039537. [Crossref] [PubMed]
49. Anagnostopoulos AK, Gaitanis A, Gkiozos I, et al. Radiomics/Radiogenomics in Lung Cancer: Basic Principles and Initial Clinical Results. Cancers (Basel) 2022;14:1657. [Crossref] [PubMed]
50. Chen M, Copley SJ, Viola P, et al. Radiomics and artificial intelligence for precision medicine in lung cancer treatment. Semin Cancer Biol 2023;93:97-113. [Crossref] [PubMed]
51. Hu Q, Li K, Yang C, et al. The role of artificial intelligence based on PET/CT radiomics in NSCLC: Disease management, opportunities, and challenges. Front Oncol 2023;13:1133164. [Crossref] [PubMed]
52. Cherukuri SP, Kaur A, Goyal B, et al. Artificial Intelligence-Enhanced Liquid Biopsy and Radiomics in Early-Stage Lung Cancer Detection: A Precision Oncology Paradigm. Cancers (Basel) 2025;17:3165. [Crossref] [PubMed]
53. Albahri AS, Duhaim AM, Fadhel MA, et al. A systematic review of trustworthy and explainable artificial intelligence in healthcare: Assessment of quality, bias risk, and data fusion. Information Fusion 2023;96:156-91.
54. Lopes S, Mascarenhas M, Fonseca J, et al. Unveiling the Algorithm: The Role of Explainable Artificial Intelligence in Modern Surgery. Healthcare (Basel) 2025;13:3208. [Crossref] [PubMed]
55. Satapathy P, Hermis AH, Rustagi S, et al. Artificial intelligence in surgical education and training: opportunities, challenges, and ethical considerations - correspondence. Int J Surg 2023;109:1543-4. [Crossref] [PubMed]
56. Déniz C, Marcè J, Macia I, et al. Artificial intelligence in thoracic surgery consultations: evaluating the concordance between a large language model and expert clinical decisions. Front Digit Health 2025;7:1633278. [Crossref] [PubMed]
57. Arjomandi Rad A, Vardanyan R, Athanasiou T, et al. The ethical considerations of integrating artificial intelligence into surgery: a review. Interdiscip Cardiovasc Thorac Surg 2025;40:ivae192. [Crossref] [PubMed]