State-of-the-art artificial intelligence methods for pre-operative planning of cardiothoracic surgery and interventions: a narrative review

Introduction

The field of cardiothoracic surgery (CTS) encompasses the surgical management of diseases, abnormalities, and trauma within the thoracic cavity (1). In 2021, approximately 900,000 cardiothoracic procedures were performed in the United States, addressing conditions such as coronary artery disease (CAD), heart valve disease, pathology of the large intrathoracic vessels, thoracic wall pathology, mediastinal tumors, and lung cancer (2,3). The outcome of cardiothoracic procedures depends on various factors, including patient characteristics (age, comorbidities) and (peri)operative factors, such as the occurrence of complications, surgical team experience, and preoperative surgical planning (4,5).

Over recent decades, surgical procedures have become more complex, partly due to the rise of minimally invasive (robotic) surgery (MIS). While MIS allows for surgery through smaller incisions with reduced surgical trauma, procedures like video-assisted thoracic surgery (VATS) or robotic-assisted thoracic surgery (RATS) are often more complex due to the absence of a direct view of anatomical structures. Consequently, advanced (imaging) preparations are imperative, requiring a detailed understanding of patient anatomy (6,7). Novel technologies, including three-dimensional (3D) visualization, virtual and augmented reality (VR, AR) and 3D printing, have emerged as integral components of the standard workflow in cardiothoracic surgical planning, potentially enhancing anatomical awareness and intraoperative orientation, leading to improved surgical outcomes (8-10).

The use of artificial intelligence (AI) offers a promising opportunity to support cardiothoracic surgeons in disease detection, personalized treatment planning, and predicting surgical outcomes (11). AI refers to the field of computer science that focuses on the development of algorithms commonly associated with human intelligence, critical thinking and cognitive activities, with machine learning (ML) and deep learning (DL) as specific subfields (12). ML focuses on developing algorithms and models that enable learning from data to recognize patterns and make predictions for prospective events, while DL, a subfield of ML, employs deep artificial neural networks with multiple layers to automatically learn hierarchical representations of data, enabling the extraction of patterns and features (13,14).

AI can be used for several tasks in preoperative planning of (complex) cardiothoracic surgeries (11). The first task is image segmentation, the process of delineating structures within medical images such as X-ray images, computed tomography (CT) images, and magnetic resonance imaging (MRI) images. Visualizing segmented anatomical structures alongside imaging data can provide an improved understanding of the patient anatomy (11). Secondly, AI can support in diagnosis and surgical decision making (11). For example, a retrospective study in elective cardiac surgery cases showed that an ML model was more accurate in predicting mortality than the Euroscore II calculator, a tool used in cardiac surgery to estimate the risk of mortality for patients undergoing cardiac surgery (15). AI can automate several important preoperative steps, such as detection, classification, and measurements. For example, in tumor and lesion detection, several studies developed ML or DL algorithms to detect and classify lung nodules with high sensitivity and specificity (16-18). Measurements of, for instance, aorta diameter, radius of aortic curvature, mitral valve annulus size, and aortic valve annulus size are usually performed manually by a radiologist or surgeon and are time-consuming. With the use of AI, these measurements can be performed with the same accuracy but faster (19).

Previous reviews have focused on the initiatives using AI in CTS (11), the role and future of AI in CTS (20) and the impact of AI for cardiothoracic surgeons (21). However, to the best of our knowledge, there is no review on AI applications for the specific purpose of preoperative planning of cardiothoracic surgical procedures. Consequently, the aim of this study is to review and provide an overview of state-of-the-art AI techniques in preoperative planning of cardiothoracic surgeries and interventions, encompassing aortic (valve) surgery and intervention [including transcatheter aortic valve implantation (TAVI) planning], mitral valve surgery and intervention planning, other cardiac surgery, and lung, thoracic wall and mediastinum surgery. We present this article in accordance with the Narrative Review reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-24-1793/rc).

Methods

Search strategy

For this literature review, we used the following keywords in the search: CTS, AI, ML, DL, and preoperative planning. These keywords resulted in a literature search with two search strings conducted on the PubMed database. All articles published before August 13th, 2024, were considered for inclusion in this review. After the independent screening of articles based on title and abstract by two researchers (Q.J.M., A.H.S.), a thorough screening of the full articles was conducted. Subsequently, a full-text analysis was carried out on the remaining articles. All articles were discussed with a second observer, and in case of doubt, a discussion was held to decide on the inclusion of the article. In addition to the articles identified through the PubMed search, four additional articles [Saitta et al. (22), Huang et al. (23), Chang et al. (24) and Salati et al. (25)] were considered relevant and included following a comprehensive assessment of the full text and references analysis (Table 1).

Exclusion criteria

The exclusion criteria encompassed: (I) studies involving animals, cadavers, or phantom studies; (II) studies focusing on intraoperative or postoperative planning and evaluation; (III) studies not related to CTS; (IV) studies lacking AI components; (V) articles not written in English; (VI) unavailability of articles; (VII) duplicates; and (VIII) studies scoring less than 3 out of 4 in the quality assessment (Table 1).

Quality assessment

A quality assessment was conducted utilizing the Centre for Evidence-Based Medicine (CEBM) worksheet (26). This worksheet was used to address the following questions for quality assessment: (I) Does this study address a clearly focused question? (II) Did the study use valid methods to address this question? (III) Are the valid results of this study important? (IV) Are these valid, important results applicable to other populations? Articles with a minimum score of 3 out of 4 were included for further analysis.

Data extraction

The AI approaches used in the preoperative planning of CTS procedures were obtained from the articles. Additionally, extracted data included image modality, target anatomical tissue, target (surgical) procedure, AI task, and evaluation score for the AI method (Table 1).

Results

Study inclusion and characteristics

A total of 525 articles were initially identified through the two search strings. Following title and abstract screening, 411 articles were excluded based on the following criteria: absence of AI (n=288), unrelated to CTS (n=50), absence of preoperative planning (n=26), lack of human data (n=7), unavailability of articles (n=30), and duplicates (n=10). Subsequently, after a thorough screening of the full text, an additional 74 articles were excluded (absence of AI: n=12, absence of preoperative planning: n=44, unrelated to CTS: n=4, unavailability of articles: n=12, and non-English language: n=2). Upon full-text analysis, 12 articles were further excluded due to the absence of AI. Four articles were additionally included based on their relevance identified through cross-referencing. This process led to the inclusion of 32 articles for this review. The details of the literature search are presented in Figures 1,2. The articles were categorized based on the target thoracic surgical application into the following categories: (I) aortic (valve) surgery/intervention (including TAVI planning); (II) mitral valve surgery/intervention planning; (III) other cardiac surgery; and (IV) lung, thoracic wall and mediastinum surgery (Figure 2).

Figure 1 Article selection and PRISMA flow diagram. AI, artificial intelligence; CTS, Cardiothoracic surgery; preop, preoperative planning.

Figure 2 Overview of the included literature. (a descriptive caption). (A) Pie chart of the included papers based on the predetermined categories. (B) Bar chart of the number of publications included for this review based on year of publication.

Aortic (valve) surgery/intervention planning

For selected patients diagnosed with aortic stenosis (AS) who are at high risk and require an intervention, TAVI has become an alternative to surgical aortic valve replacement (SAVR) (27). Efficient preoperative planning of TAVI prosthesis sizing can enhance workflow efficiency.

TAVI planning involves anatomical measurements of the aortic annulus perimeter, aortic annulus area, aortic valve annulus size, location and orientation of the valve annular plane, and coronary ostia. Astudillo et al. (28) developed a DL method for automatic measurements from multidetector computed tomography (MDCT) images, validated through an interobserver variability study. Al et al. (29), Astudillo et al. (30), and Theriault-Lauzier et al. (31) also introduced AI approaches for automatic detection of aortic landmarks. Al et al. reported increased accuracy by comparing their proposed ML algorithm with manual ground truth annotations (29). Astudillo et al. demonstrated similar accuracy for DL model predictions of these landmarks compared to ground truth annotations. Additionally, a high correlation between automated and manually measured coronary ostia heights was observed (30). An accuracy of 84.6% in determining the location and orientation of the valve annular plane, based on aortic valve annulus segmentation, was achieved by Theriault-Lauzier et al. (31). A Dice score of 0.5±0.2 was observed for segmentation of the aortic annulus plane by Cho et al. (32). Santaló-Corcoy et al. proposed a fully automated DL method for TAVI planning (33). Their method can extract 22 relevant measurements [such as annulus area, annulus diameter, and left ventricular outflow tract (LVOT) diameters] from the aortic valvular complex, which are essential for TAVI planning (33). Using a DL algorithm, Wang et al. efficiently performed pre-TAVI assessments using CT scans with an accuracy comparable to senior observers (34). The AI methods in these studies resulted in reduced processing time for landmark detection and measurements to below 1 second (30), 12 milliseconds (29), 2 minutes (33) and 0.86±0.21 min (34) compared to manual annotation (5–10 minutes).

Saffar et al. evaluated the accuracy of a DL-based algorithm designed to detect thoracic aortic calcifications in chest CT scans and their utility in cardiovascular surgery planning (35). The study assessed the performance of the algorithm in automatically identifying and quantifying aortic calcifications, which are critical for risk assessment and surgical decision-making. The findings demonstrated that the algorithm provides highly accurate detection and quantification, aligning well with expert radiologist assessments.

Segmentation of anatomical structures such as the aortic valve, aorta, left ventricle aortic root, and LVOT can be beneficial in preoperative planning. Queiros et al. performed a fully automated segmentation of the aortic valve from 3D transesophageal echocardiogram (TEE) images (36). In three other studies, DL approaches were used for automatic segmentation of structures such as the aortic valve, aorta, left ventricle aortic root, and LVOT. Three studies showed average Dice scores of 0.95 for aorta and left ventricle (37), 0.94 for aorta and aortic valve (38), and 0.93 for aortic root, aortic annulus, and sinotubular junction (39) using their proposed DL models. Saitta et al. also presented a fully automated DL pipeline for thoracic aorta geometric analysis and planning for thoracic endovascular aortic repair (TEVAR) from CT images (22). With the DL model, automatic thoracic (aneurysmatic) aorta segmentation (Dice score of 0.95), detection of proximal landing zones, and quantification of geometric measurements (such as aortic arch centerline and radius of aortic curvature) can be performed. These studies indicated decreased segmentation times compared to manual segmentation, which usually requires 30–60 minutes [Queiros et al. (36): 12 seconds, Krüger et al. (38): 30 seconds, Saitta et al. (39): 45 seconds, Saitta et al. (22): 7 minutes]. As mentioned earlier, segmentation can be useful in preoperative planning, where anatomical 3D reconstruction can potentially improve surgical outcome and is even recommended by the European Society of Thoracic Surgery (ESTS) (40).

The integration of AI in surgical planning software can be useful in clinical practice. The papers by Boninsegna et al. (41) and Toggweiler et al. (42) both explore the use of AI in improving medical imaging and procedural planning. Boninsegna et al. (41) focused on the potential of AI to enhance the reporting of CT angiography (CTA) prior to endovascular procedures, improving accuracy, speed, and consistency in image analysis. Toggweiler et al. (42) discussed a fully automated AI-driven software designed to streamline transcatheter aortic valve replacement (TAVR) planning, automating key steps like image analysis and valve sizing to reduce errors and optimize workflow. These studies highlight the growing role of AI in automating and enhancing the planning of complex cardiovascular interventions, improving both accuracy and efficiency.

Other important applications of AI include preoperative evaluation, risk management, and identification of distinct phenotypes of AS severity. A novel approach utilizing ML techniques was introduced by Sengupta et al. to identify and categorize distinct phenotypes of AS severity (43). While traditional methods rely on manual measurements and subjective interpretation of echocardiographic images, the ML framework proposed by Sengupta et al. can categorize AS severity into different phenotypes based on important clinical parameters such as valve area and transvalvular pressure gradient from echo images. The ML framework successfully identifies distinct AS severity phenotypes with a high degree of accuracy and provides insights into the underlying mechanisms and characteristics of each phenotype (43).

Publications on the application of AI in aortic (valve) surgery and interventions are presented in Table 2.

Mitral valve surgery/intervention

In percutaneous mitral valve procedures, such as transcatheter mitral valve replacement (TMVR), preoperative planning involves anatomical analysis and LVOT assessment to mitigate the risk of LVOT obstruction by the prosthesis. Jeganathan et al. (44) used AI software (eSie Valve software) to measure six parameters: (I) mitral annulus anterolateral posteromedial diameter, (II) mitral annulus anteroposterior diameter, (III) mitral annular area, (IV) mitral annulus non-planarity angle, (V) mitral annulus total perimeter, and (VI) anterior and posterior leaflet area. Utilizing the AI software for 3D transesophageal echocardiography image analysis demonstrated comparable accuracy to human observers, with no significant differences. Notably, the use of this AI software resulted in reduced time and effort for surgeons during preoperative planning (44). Astudillo et al. employed DL models for the automatic detection and annotation of the mitral valve annulus (45). The DL models facilitated segmentation and measurement of mitral valve annulus parameters (2D perimeter, trigone-to-trigone distance, septal-to-lateral distance and commissure-to-commissure distance) for anatomical analysis crucial for accurate device size determination. The study revealed a high correlation between automatic and manual measurements, underscoring the reproducibility and accuracy achievable with DL in obtaining mitral valve measurements. Additionally, the automated DL method significantly expedited the measuring process, taking only one second compared to the 25 minutes required for manual measurements. Publications on the application of AI for mitral valve surgery and interventions are presented in Table 3.

Other cardiac surgery

In the correction of total anomalous pulmonary venous connection, precise segmentation and visualization of the pulmonary vein (PV) and left atrium (LA) play a crucial role and can provide added value to the surgeon performing the procedure. Li et al. trained a DL algorithm for PV and LA segmentation utilizing low-dose CT images. The DL model achieved a mean Dice score of 0.80 (PV) and 0.83 (LA). Workflow efficiency was substantially improved, reducing the processing time for segmentation per case from 2–3 hours (manual) to 400 milliseconds (DL) per case (46).

As previously highlighted, risk stratification and disease prediction are also applications where AI can play a role. Baskaran et al. conducted a study utilizing ML techniques to assess the predictive value of imaging and clinical variables in determining the presence of obstructive CAD and the need for revascularization procedures (47). These algorithms were compared with the standard CAD2 score, which demonstrates the accuracy for predicting CAD in low-risk populations and includes cardiovascular risk factors such as age, sex and, hypertension. The ML algorithms integrate imaging variables, such as coronary artery calcium score and coronary CTA findings, with clinical variables encompassing demographic information, medical history, and laboratory results. The ML algorithm outperformed the CAD2 score for predicting obstructive CAD. This suggests that the ML approach can serve as a non-invasive diagnostic tool for determining CAD (47). Publications on the application of AI for other cardiac surgeries are presented in Table 3.

Lung, thoracic wall and mediastinal surgery

A detailed and comprehensive anatomical knowledge of the lung, arteries, veins, and bronchi is essential for effective planning and execution of lung surgery. In the context of lung and airway surgery, several studies have explored the application of AI for image segmentation. Meng et al. proposed an ML method for airway segmentation, achieving segmentation of 79.1% of patients’ airways compared to manual segmentation by an expert, considered as the gold standard (48). Kockelkorn et al. developed an interactive lung segmentation method based on an ML algorithm, which successfully segmented lung parenchyma within 2 minutes (49).

For lung sparing segmentectomy, where detailed anatomical knowledge and preoperative planning are crucial for accurate resections, surgical margins, and patients’ outcome, three studies developed fully automated DL algorithms for segmenting arteries, veins and bronchi based on CT scans (50-52). Chen et al. reported an overall Dice of 0.70 compared to manual segmentation (50). The automatic detection of pulmonary vessels and bronchi was 85%, compared to 80% for manual detection, and the accuracy of vessel classification using the AI algorithm was 80%, compared to 95% for manual classification. Li et al. demonstrated similar accuracy to manual segmentation for segmental bronchi and outperformed manual segmentation in segmental arteries and segmental veins (51). Chen et al. achieved an overall accuracy of 82.8%, compared to 78.8% and 77.0% for manual segmentation by two surgeons (52). The accuracy for segmental artery and lobular vein did not show any significant differences between manual and automatic segmentation. Pan et al. achieved an average accuracy of 90.1% on an external test set, and 96.2% on their own local dataset (53). A decrease in reconstruction time was observed by Chen et al. (AI =2 minutes, manual =30 minutes) and Li et al. (6.8 vs. 21 minutes).

The usefulness of AI in automatic preoperative planning was demonstrated by Sadeghi et al., who combined AI and VR for preoperative planning of segmentectomy procedures (54). Using a DL algorithm, segmentation of the arteries, veins, airways, and lung segments of the lung lobe containing a tumor was obtained. By loading those segments into a virtual environment and using VR, a 3D reconstruction of the patient could be displayed and used for accurate preoperative planning. A recent study demonstrated a significant change in surgical plan when the AI-VR approach was used compared to conventional preoperative planning with 2D CT scans (55).

In addition to segmentation, Ding et al. introduced two DL approaches (Lung-DL model and Dense model) to assess invasiveness and predict the presence of the micropapillary pattern within lung adenocarcinoma using CT scans (56). Their DL algorithms successfully classified the invasiveness of pulmonary adenocarcinomas and can predict micropapillary patterns. These algorithms can increase planning efficiency. Mayoral et al. introduced an ML algorithm to differentiate malignant from benign tumors in the mediastinum based on CT images and, secondly, to differentiate thymomas from thymic carcinomas (57). The model using both conventional features (such as tumor size and shape) and radiomic features achieved the highest diagnostic performance, compared to conventional-only and radiomic-only models. Similarly, differentiation between thymoma and thymic carcinomas achieved the highest diagnostic performance in the combined conventional and radiomic features model, indicating that combining conventional features with radiomic features from CT images in an ML model can be useful for identifying diseases of the mediastinum (57).

AI can also be used to predict postoperative cardiopulmonary complications after lung resections. Huang et al. (23), Chang et al. (24), and Salati et al. (25) used ML algorithms to predict postoperative complications based on medical records and patient characteristics such as age, gender, smoking history, and comorbidities, with the highest area under the curve (AUC) of 0.767, 0.912 and 0.75, respectively. All three ML methods have the potential to enhance the counseling process and preoperative management of patients undergoing lung resection.

Publications on the application of AI for lung, thoracic wall and mediastinal surgery are presented in Table 4.

Discussion

The role and application of AI in CTS have witnessed a significant increase in recent years. AI holds potential in pattern recognition, decision-making processes, and supporting patient-specific preoperative planning for complex interventions. This review aims to provide an overview of state-of-the-art AI methods applicable to preoperative planning in CTS and interventions.

In TAVI procedures, AI methods play a crucial role in efficiently achieving reproducible and accurate device sizing, as well as in pre- and postoperative risk assessment. Consequently, AI methods that facilitate this can be helpful for surgeons and interventional cardiologists in preoperative planning of TAVI (58). These methods, as presented in Table 2, are instrumental for optimal TAVI planning. Although valve-related parameters (e.g. coronary ostia heights, aortic valve annulus diameter) are important for device sizing in TAVI planning, other factors can play an important role in defining treatment outcome. For example, patient-specific computational models to predict aortic regurgitation or conduction abnormalities after TAVI have been proposed by several groups (58,59). In these studies, the authors demonstrated that computer simulations of a TAVI procedure might be beneficial in predicting the amount of aortic regurgitation or conduction abnormalities after device implantation. While AI has proven to be valuable in preoperative measurements of relevant aortic valve parameters, it may also be helpful in automating and refining TAVI procedure planning. This can be achieved through both DL and ML techniques. Naturally, these applications could also be used for transcatheter interventions on other valves (e.g., TMVR).

Segmentation of important structures is a useful application of AI in preoperative planning, facilitating 3D reconstruction of patients’ anatomy from CT or MRI images (20). Various DL or ML algorithms have been proposed to segment pulmonary arteries, veins, bronchi, tumors, lobes, and segments (53-60). However, not many of these algorithms are available for routine clinical use in lung surgery planning. As can be seen in the results section, our group has integrated an existing AI tool into pulmonary surgery planning, combining AI with VR technology for enhanced understanding of complex lung and heart anatomy (54,55). The use of labeled structures on top of the regular imaging modalities and specific visualization tools, such as PulmoVR (Surgical Reality, Nieuw Vennep, The Netherlands), in combination with the segmented structures, can help surgeons better understand complex lung and heart anatomy, improving their knowledge of the patient-specific preoperative anatomy. These segmentations contribute to simulations of resections, aiding in predictions of postoperative reduction in lung volume, which can also be helpful in risk stratification.

Preoperative planning with good understanding of patient-specific anatomy is a very important step for a successful execution of complex procedures. For many of the reported articles in this review, the use of AI can achieve similar accuracy to manually obtained measurements for tasks such as parameter calculation, image segmentation, prediction of risks, and clinical decision making. As described, the use of AI in preoperative planning can offer several advantages. As shown in Tables 2-4, an overall decrease in preoperative planning time can be achieved using AI in the regular workflow. Compared with manual annotation, which usually can take up to several hours, AI can reduce this processing time to a couple of minutes or even seconds. Secondly, AI can support surgeons and clinicians in surgical decision-making preoperatively. By using AI to measure relevant parameters in preoperative planning, decisions can be made in a more efficient manner.

Despite the considerable advancements in AI research, the integration into the daily workflow of cardiothoracic surgeons remains limited. AI is described as a supportive tool, not a replacement for healthcare professionals such as interventional cardiologists or cardiothoracic surgeons (22,28,50,51). In medical decision making, the surgeon remains responsible, despite the use of techniques to provide extra information on patient anatomy. Several factors need to be taken into account in surgical decision making, such as medical history, complication risks, prognosis, and mental state of the patient (11,20). All of these factors are important in treatment planning for patients and require surgical experience, which is not available in the AI model (yet) (61).

One of the primary concerns of AI in healthcare is the representativeness, or generalizability, and robustness of AI models (62,63). While most articles in this review provide a comprehensive description of their dataset, the majority of the patients included to train their AI model are predominantly obtained from their respective institutions. As a result, the generalizability of the AI models to populations not adequately represented in the training data, may be compromised. Furthermore, the enhancement of diversity and inclusion of various conditions in the training data has been discussed in several studies. Lastly, the availability of publicly available AI algorithms is limited, as most of the algorithms are developed for study purposes in a single center. Therefore, in many studies, expanding the dataset, including multiple conditions and utilizing multicenter data are described as areas for improvement in the study.

The integration of AI into CTS holds significant potential for enhancing clinical outcomes while addressing cost concerns. AI-driven predictive analytics improve preoperative planning by enhancing patient risk stratification and optimizing resource use, reducing unnecessary tests and hospital stays (64). Although the initial implementation costs are high, long-term financial benefits include improved efficiency, fewer surgical errors, and better patient outcomes (65). Studies show that these advantages lead to gradual recovery of initial investments and sustained savings through shorter hospitalizations and fewer follow-up procedures (20,66). However, challenges such as high costs, the need for specialized training, and ethical concerns regarding data privacy and transparency remain. The variability in cost-effectiveness across different healthcare settings calls for further research and tailored strategies (67). Future studies should focus on understanding long-term cost-effectiveness and developing standardized protocols to ensure consistent improvements and cost savings in CTS. AI models have different accuracy requirements depending on the surgical procedure. For example, in lung segmentectomy planning, segmental and lobar segmentation of structures are of higher importance than peripheral small sub-branches. A relatively lower Dice score compared to a detailed manual segmentation (of peripheral anatomical branches) does not necessarily mean that the AI model is not usable for surgical planning. Therefore, adding a clinical validation method to describe the effect of AI in medical applications will clarify the usability of AI models for preoperative planning for clinicians.

A thorough analysis of the limitations of AI models in the context of CTS preoperative planning reveals several key challenges that could impact their clinical effectiveness and widespread adoption. Despite the considerable promise of AI to enhance surgical precision, and potentially improve patient outcomes, these limitations must be addressed to ensure their successful integration into clinical practice and to maximize their potential benefits (68).

A primary limitation of AI models in CTS lies in the variability of data used to train them. CTS deals with a wide range of patient characteristics, including diverse anatomies, comorbidities, and surgical conditions. AI models are often trained on specific datasets, which may not represent the full spectrum of patients seen in clinical practice (69,70). This lack of data diversity can result in models that perform poorly when applied to patient populations that differ significantly from the dataset used for training. This reduces the reliability and generalizability of AI models, particularly when dealing with complex or rare cases, and creates a risk in using AI in broader patient populations (71). As AI models are generally trained using data from controlled or homogeneous populations, they may struggle with the inherent heterogeneity found in real-world clinical environments (69).

Another significant limitation of many AI models is the “black box” nature of their decision-making process. In CTS, where the complexity of surgical decisions is high, understanding how an AI system arrives at a particular recommendation is critical (72). AI models, particularly DL systems, are often difficult to interpret, making it challenging for clinicians to fully trust their outputs (73). When AI models cannot explain why a certain recommendation is made, it becomes difficult for clinicians to evaluate the reasoning behind the decision (68). This lack of transparency raises concerns regarding accountability, especially in the event of negative patient outcomes. It is crucial that AI systems be designed with interpretability in mind, as trust and reliability in these tools are essential for their successful clinical implementation (74).

The integration of AI models into existing clinical workflows remains a significant challenge. AI systems designed to aid in preoperative planning must be seamlessly incorporated into the existing infrastructure of hospitals and surgical teams. This requires compatibility with electronic health records (EHRs), imaging systems, and other clinical technologies, which can be a complex and resource-intensive process (70,73). Integrating AI into the clinical workflow could require significant adjustments to how healthcare providers manage and interact with patient data, disrupting established practices. Additionally, AI models often demand significant computational resources, including powerful hardware and specialized software, which may not be readily available in all clinical settings (71). In resource-limited environments, this could present a barrier to adoption, limiting AI’s availability to larger hospitals or those with more advanced technological infrastructure (72).

Limitations

This review on the use of AI in preoperative planning of cardiothoracic surgeries has several limitations that could impact the robustness and generalizability of the findings. The present study is constrained by the inclusion of exclusively English-language articles, which may introduce potential bias in the selection of papers. Additionally, only the PubMed database was utilized in this review. The incorporation of multiple databases could potentially result in the inclusion of a greater number of relevant articles on this topic. The field of AI applications in CTS is still evolving, resulting in a limited number of eligible studies, which may compromise the comprehensiveness of this review. Furthermore, this review did not incorporate a meta-analysis but was rather a narrative on state-of-the-art AI technologies used for pre-operative planning in CTS. The primary rationale behind this decision was the substantial heterogeneity observed across the selected studies. Variability in methodologies and outcome measures, for example, among the included studies, introduced a level of diversity that posed challenges to achieving a homogenous integration of results. Importantly, it should be emphasized that opting against a meta-analysis does not diminish the significance of this review. Instead, the focus was placed on a thorough qualitative evaluation of individual studies. This approach allowed for a nuanced examination of the unique contributions of each study, providing valuable insights into the diverse facets of this research topic. Recognizing and addressing these limitations within the review process will contribute to a more nuanced interpretation of the findings.

Conclusions

We have presented a comprehensive overview of AI algorithms designed to facilitate preoperative planning in CTS. The application of AI in various cardiothoracic interventions enables automated measurements primarily based on prevalent imaging modalities. The incorporation of AI in preoperative workflows yields a substantial reduction in processing time and planning duration, thereby potentially enhancing overall efficiency. Furthermore, AI serves as a valuable tool in supporting clinical and surgical decision-making processes. The prospective role of AI in CTS planning appears promising, with the potential to contribute significantly to tasks such as image segmentation, measurement, and calculation automation, prognosis prediction, and guidance in surgical decision-making. However, AI will most likely not replace clinicians but can serve as a supportive tool. More research is needed to validate the use of certain AI models in clinical practice.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-24-1793/rc

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-24-1793/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-24-1793/coif). Q.J.M. is a part-time employee of Surgical Reality. A.H.S. is a co-inventor of Surgical Reality. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are 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. Hennessy SA, Lau CL. Cardiothoracic surgery. Am J Surg 2011;202:357-9. [Crossref] [PubMed]
2. Bowdish ME, D'Agostino RS, Thourani VH, et al. STS Adult Cardiac Surgery Database: 2021 Update on Outcomes, Quality, and Research. Ann Thorac Surg 2021;111:1770-80. [Crossref] [PubMed]
3. Doenst T, Caldonazo T, Schneider U, et al. Cardiac Surgery 2020 Reviewed. Thorac Cardiovasc Surg 2021; Epub ahead of print. [Crossref]
4. Bapat V, Allen D, Young C, et al. Survival and quality of life after cardiac surgery complicated by prolonged intensive care. J Card Surg 2005;20:212-7. [Crossref] [PubMed]
5. Ghanta RK, Shekar PS, McGurk S, et al. Long-term survival and quality of life justify cardiac surgery in the very elderly patient. Ann Thorac Surg 2011;92:851-7. [Crossref] [PubMed]
6. Totaro P, Carlini S, Pozzi M, et al. Minimally invasive approach for complex cardiac surgery procedures. Ann Thorac Surg 2009;88:462-6; discussion 467. [Crossref] [PubMed]
7. Anderson RH, Wilcox BR. Understanding cardiac anatomy: the prerequisite for optimal cardiac surgery. Ann Thorac Surg 1995;59:1366-75. [Crossref] [PubMed]
8. Wellens F, Goldstein JBelgian Association for Cardiac Surgery. New developments in cardiac surgery. Acta Cardiol 2004;59:35-8.
9. Onugha O, Ivey R, McKenna R. Novel Techniques and Approaches to Minimally Invasive Thoracic Surgery. Surg Technol Int 2017;30:231-5.
10. Arjomandi Rad A, Vardanyan R, Thavarajasingam SG, et al. Extended, virtual and augmented reality in thoracic surgery: a systematic review. Interact Cardiovasc Thorac Surg 2022;34:201-11. [Crossref] [PubMed]
11. Dias RD, Shah JA, Zenati MA. Artificial intelligence in cardiothoracic surgery. Minerva Cardioangiol 2020;68:532-8. [Crossref] [PubMed]
12. Amisha Malik P. Overview of artificial intelligence in medicine. J Family Med Prim Care 2019;8:2328-31. [Crossref] [PubMed]
13. Badillo S, Banfai B, Birzele F, et al. An Introduction to Machine Learning. Clin Pharmacol Ther 2020;107:871-85. [Crossref] [PubMed]
14. Choi RY, Coyner AS, Kalpathy-Cramer J, et al. Introduction to Machine Learning, Neural Networks, and Deep Learning. Transl Vis Sci Technol 2020;9:14. [Crossref] [PubMed]
15. Allyn J, Allou N, Augustin P, et al. A Comparison of a Machine Learning Model with EuroSCORE II in Predicting Mortality after Elective Cardiac Surgery: A Decision Curve Analysis. PLoS One 2017;12:e0169772. [Crossref] [PubMed]
16. Liu X, Zhou H, Hu Z, et al. Clinical Application of Artificial Intelligence Recognition Technology in the Diagnosis of Stage T1 Lung Cancer. Zhongguo Fei Ai Za Zhi 2019;22:319-23. [Crossref] [PubMed]
17. Zhang C, Sun X, Dang K, et al. Toward an Expert Level of Lung Cancer Detection and Classification Using a Deep Convolutional Neural Network. Oncologist 2019;24:1159-65. [Crossref] [PubMed]
18. Zhao L, Bai CX, Zhu Y. Diagnostic value of artificial intelligence in early-stage lung cancer. Chin Med J (Engl) 2020;133:503-4. [Crossref] [PubMed]
19. Orfanoudaki A, Dearani JA, Shahian DM, et al. Improving Quality in Cardiothoracic Surgery: Exploiting the Untapped Potential of Machine Learning. Ann Thorac Surg 2022;114:1995-2000. [Crossref] [PubMed]
20. Mumtaz H, Saqib M, Ansar F, et al. The future of Cardiothoracic surgery in Artificial intelligence. Ann Med Surg (Lond) 2022;80:104251. [Crossref] [PubMed]
21. Nedadur R, Wang B, Yanagawa B. The cardiac surgeon's guide to artificial intelligence. Curr Opin Cardiol 2021;36:637-43. [Crossref] [PubMed]
22. Saitta S, Sturla F, Caimi A, et al. A Deep Learning-Based and Fully Automated Pipeline for Thoracic Aorta Geometric Analysis and Planning for Endovascular Repair from Computed Tomography. J Digit Imaging 2022;35:226-39. [Crossref] [PubMed]
23. 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]
24. Chang YJ, Hung KC, Wang LK, et al. A Real-Time Artificial Intelligence-Assisted System to Predict Weaning from Ventilator Immediately after Lung Resection Surgery. Int J Environ Res Public Health 2021;18:2713. [Crossref] [PubMed]
25. Salati M, Migliorelli L, Moccia S, et al. A Machine Learning Approach for Postoperative Outcome Prediction: Surgical Data Science Application in a Thoracic Surgery Setting. World J Surg 2021;45:1585-94. [Crossref] [PubMed]
26. Centre for Evidence-Based Medicine (CEBM), University of Oxford. Critical appraisal tools (2021, June 14). Retrieved August 3, 2022. Available online: https://www.cebm.ox.ac.uk/
27. Ben-Dor I, Dvir D, Barbash IM, et al. Outcomes of patients with severe aortic stenosis at high surgical risk evaluated in a trial of transcatheter aortic valve implantation. Am J Cardiol 2012;110:1008-14. [Crossref] [PubMed]
28. Astudillo P, Mortier P, Bosmans J, et al. Enabling Automated Device Size Selection for Transcatheter Aortic Valve Implantation. J Interv Cardiol 2019;2019:3591314. [Crossref] [PubMed]
29. Al WA, Jung HY, Yun ID, et al. Automatic aortic valve landmark localization in coronary CT angiography using colonial walk. PLoS One 2018;13:e0200317. [Crossref] [PubMed]
30. Astudillo P, Mortier P, Bosmans J, et al. Automatic Detection of the Aortic Annular Plane and Coronary Ostia from Multidetector Computed Tomography. J Interv Cardiol 2020;2020:9843275. [Crossref] [PubMed]
31. Theriault-Lauzier P, Alsosaimi H, Mousavi N, et al. Recursive multiresolution convolutional neural networks for 3D aortic valve annulus planimetry. Int J Comput Assist Radiol Surg 2020;15:577-88. [Crossref] [PubMed]
32. Cho Y, Park S, Hwang SH, et al. Aortic Annulus Detection Based on Deep Learning for Transcatheter Aortic Valve Replacement Using Cardiac Computed Tomography. J Korean Med Sci 2023;38:e306. [Crossref] [PubMed]
33. Santaló-Corcoy M, Corbin D, Tastet O, et al. TAVI-PREP: A Deep Learning-Based Tool for Automated Measurements Extraction in TAVI Planning. Diagnostics (Basel) 2023;13:3181. [Crossref] [PubMed]
34. Wang M, Niu G, Chen Y, et al. Development and validation of a deep learning-based fully automated algorithm for pre-TAVR CT assessment of the aortic valvular complex and detection of anatomical risk factors: a retrospective, multicentre study. EBioMedicine 2023;96:104794. [Crossref] [PubMed]
35. Saffar R, Sperl JI, Berger T, et al. Accuracy of a deep learning-based algorithm for the detection of thoracic aortic calcifications in chest computed tomography and cardiovascular surgery planning. Eur J Cardiothorac Surg 2024;65:ezae219. [Crossref] [PubMed]
36. Queiros S, Papachristidis A, Morais P, et al. Fully Automatic 3-D-TEE Segmentation for the Planning of Transcatheter Aortic Valve Implantation. IEEE Trans Biomed Eng 2017;64:1711-20. [Crossref] [PubMed]
37. Zlahoda-Huzior A, Stanuch M, Witowski J, et al. Automatic aorta and left ventricle segmentation for TAVI procedure planning using convolutional neural networks. Annu Int Conf IEEE Eng Med Biol Soc 2019;2019:2777-80. [Crossref] [PubMed]
38. Krüger N, Meyer A, Tautz L, et al. Cascaded neural network-based CT image processing for aortic root analysis. Int J Comput Assist Radiol Surg 2022;17:507-19. [Crossref] [PubMed]
39. Saitta S, Sturla F, Gorla R, et al. A CT-based deep learning system for automatic assessment of aortic root morphology for TAVI planning. Comput Biol Med 2023;163:107147. [Crossref] [PubMed]
40. Brunelli A, Decaluwe H, Gonzalez M, et al. European Society of Thoracic Surgeons expert consensus recommendations on technical standards of segmentectomy for primary lung cancer. Eur J Cardiothorac Surg 2023;63:ezad224. [Crossref] [PubMed]
41. Boninsegna E, Piffer S, Simonini E, et al. CT angiography prior to endovascular procedures: can artificial intelligence improve reporting? Phys Eng Sci Med 2024;47:643-9. [Crossref] [PubMed]
42. Toggweiler S, Wyler von Ballmoos MC, Moccetti F, et al. A fully automated artificial intelligence-driven software for planning of transcatheter aortic valve replacement. Cardiovasc Revasc Med 2024;65:25-31. [Crossref] [PubMed]
43. Sengupta PP, Shrestha S, Kagiyama N, et al. A Machine-Learning Framework to Identify Distinct Phenotypes of Aortic Stenosis Severity. JACC Cardiovasc Imaging 2021;14:1707-20. [Crossref] [PubMed]
44. Jeganathan J, Knio Z, Amador Y, et al. Artificial intelligence in mitral valve analysis. Ann Card Anaesth 2017;20:129-34. [Crossref] [PubMed]
45. Astudillo P, De Beule M, Dambre J, et al. Towards safe and efficient preoperative planning of transcatheter mitral valve interventions. Morphologie 2019;103:139-47. [Crossref] [PubMed]
46. Li J, Chen H, Zhu F, et al. Automatic Pulmonary Vein and Left Atrium Segmentation for TAPVC Preoperative Evaluation Using V-Net with Grouped Attention. Annu Int Conf IEEE Eng Med Biol Soc 2020;2020:1207-10. [Crossref] [PubMed]
47. Baskaran L, Ying X, Xu Z, et al. Machine learning insight into the role of imaging and clinical variables for the prediction of obstructive coronary artery disease and revascularization: An exploratory analysis of the CONSERVE study. PLoS One 2020;15:e0233791. [Crossref] [PubMed]
48. Meng Q, Kitasaka T, Nimura Y, et al. Automatic segmentation of airway tree based on local intensity filter and machine learning technique in 3D chest CT volume. Int J Comput Assist Radiol Surg 2017;12:245-61. [Crossref] [PubMed]
49. Kockelkorn TT, Schaefer-Prokop CM, Bozovic G, et al. Interactive lung segmentation in abnormal human and animal chest CT scans. Med Phys 2014;41:081915. [Crossref] [PubMed]
50. Chen X, Wang Z, Qi Q, et al. A fully automated noncontrast CT 3-D reconstruction algorithm enabled accurate anatomical demonstration for lung segmentectomy. Thorac Cancer 2022;13:795-803. [Crossref] [PubMed]
51. Li X, Zhang S, Luo X, et al. Accuracy and efficiency of an artificial intelligence-based pulmonary broncho-vascular three-dimensional reconstruction system supporting thoracic surgery: retrospective and prospective validation study. EBioMedicine 2023;87:104422. [Crossref] [PubMed]
52. Chen X, Xu H, Qi Q, et al. AI-based chest CT semantic segmentation algorithm enables semi-automated lung cancer surgery planning by recognizing anatomical variants of pulmonary vessels. Front Oncol 2022;12:1021084. [Crossref] [PubMed]
53. Pan L, Yan X, Zheng Y, et al. Automatic pulmonary artery-vein separation in CT images using a twin-pipe network and topology reconstruction. PeerJ Comput Sci 2023;9:e1537. [Crossref] [PubMed]
54. Sadeghi AH, Maat APWM, Taverne YJHJ, et al. Virtual reality and artificial intelligence for 3-dimensional planning of lung segmentectomies. JTCVS Tech 2021;7:309-21. [Crossref] [PubMed]
55. Bakhuis W, Sadeghi AH, Moes I, et al. Essential Surgical Plan Modifications After Virtual Reality Planning in 50 Consecutive Segmentectomies. Ann Thorac Surg 2023;115:1247-55. [Crossref] [PubMed]
56. Ding H, Xia W, Zhang L, et al. CT-Based Deep Learning Model for Invasiveness Classification and Micropapillary Pattern Prediction Within Lung Adenocarcinoma. Front Oncol 2020;10:1186. [Crossref] [PubMed]
57. Mayoral M, Pagano AM, Araujo-Filho JAB, et al. Conventional and radiomic features to predict pathology in the preoperative assessment of anterior mediastinal masses. Lung Cancer 2023;178:206-12. [Crossref] [PubMed]
58. de Jaegere P, De Santis G, Rodriguez-Olivares R, et al. Patient-Specific Computer Modeling to Predict Aortic Regurgitation After Transcatheter Aortic Valve Replacement. JACC Cardiovasc Interv 2016;9:508-12. [Crossref] [PubMed]
59. Ribeiro JM, Astudillo P, de Backer O, et al. Artificial Intelligence and Transcatheter Interventions for Structural Heart Disease: A glance at the (near) future. Trends Cardiovasc Med 2022;32:153-9. [Crossref] [PubMed]
60. Khanna A, Londhe ND, Gupta S. A Deep Attention-based U-Net for Airways Segmentation in Computed Tomography Images. Curr Med Imaging 2023;19:361-72. [Crossref] [PubMed]
61. Jiang L, Wu Z, Xu X, et al. Opportunities and challenges of artificial intelligence in the medical field: current application, emerging problems, and problem-solving strategies. J Int Med Res 2021;49:3000605211000157. [Crossref] [PubMed]
62. Faes L, Liu X, Wagner SK, et al. A Clinician's Guide to Artificial Intelligence: How to Critically Appraise Machine Learning Studies. Transl Vis Sci Technol 2020;9:7. [Crossref] [PubMed]
63. Miotto R, Wang F, Wang S, et al. Deep learning for healthcare: review, opportunities and challenges. Brief Bioinform 2018;19:1236-46. [Crossref] [PubMed]
64. Dixon D, Sattar H, Moros N, et al. Unveiling the Influence of AI Predictive Analytics on Patient Outcomes: A Comprehensive Narrative Review. Cureus 2024;16:e59954. [Crossref] [PubMed]
65. Khanna NN, Maindarkar MA, Viswanathan V, et al. Economics of Artificial Intelligence in Healthcare: Diagnosis vs. Treatment. Healthcare (Basel) 2022;10:2493. [Crossref] [PubMed]
66. Bohr A, Memarzadeh K. The rise of artificial intelligence in healthcare applications. Artificial Intelligence in Healthcare 2020;25-60.
67. Ferket BS, Oxman JM, Iribarne A, et al. Cost-effectiveness analysis in cardiac surgery: A review of its concepts and methodologies. J Thorac Cardiovasc Surg 2018;155:1671-81.e11. [Crossref] [PubMed]
68. Sulague RM, Beloy FJ, Medina JR, et al. Artificial intelligence in cardiac surgery: A systematic review. World J Surg 2024;48:2073-89. [Crossref] [PubMed]
69. Rajpurkar P, Chen E, Banerjee O, et al. AI in health and medicine. Nat Med 2022;28:31-8. [Crossref] [PubMed]
70. Vaidya YP, Shumway SJ. Artificial intelligence: The future of cardiothoracic surgery. J Thorac Cardiovasc Surg 2025;169:1265-70. [Crossref] [PubMed]
71. Constable MD, Shum HPH, Clark S. Enhancing surgical performance in cardiothoracic surgery with innovations from computer vision and artificial intelligence: a narrative review. J Cardiothorac Surg 2024;19:94. [Crossref] [PubMed]
72. Morone G, De Angelis L, Martino Cinnera A, et al. Artificial intelligence in clinical medicine: a state-of-the-art overview of systematic reviews with methodological recommendations for improved reporting. Front Digit Health 2025;7:1550731. [Crossref] [PubMed]
73. Bharati S, Mondal MRH, Podder P. A review on explainable artificial intelligence for healthcare: Why, how, and when? IEEE Trans Artif Intell 2024;5:1429-42.
74. Gómez-GonzálezEGomezEMárquez-RivasJArtificial intelligence in medicine and healthcare: A review and classification of current and near-future applications and their ethical and social impact.2020. doi: .