Artificial intelligence for real-time surgical phase recognition in minimal invasive inguinal hernia repair: a systematic review on behalf of TROGSS - the robotic global surgical society

INTRODUCTION

The revolutionary adoption of artificial intelligence (AI) in surgical applications has grown significantly in recent years, offering transformative potential across a variety of procedures. The use of AI in medicine dates back to the 1970s, with early systems assisting in bacterial detection, decision making, and imaging analysis^[1,2]. As AI models and programming advanced, their application in surgical practice became increasingly feasible. In recent literature, the use of AI has been appraised to be valuable in areas such as colorectal surgery, laparoscopic cholecystectomy, and cardiac surgery^[3,4]. However, most of these procedures utilize AI in specific aspects of their interventions, with notable effectiveness in assessing complication risks, avoiding intraoperative errors, and predicting mortality outcomes^[3-5].

An emerging field that has recently seen an increase in the incorporation of AI technologies is inguinal hernia repair (IHR), the most prevalent groin repair procedure^[6]. By training AI models using compiled videos and imaging data to identify specific steps, anatomical structures, and instruments during procedures^[7], medical teams can leverage real-time assistance to enhance decision making, improve surgical efficiency, and support education and workflow optimization^[8,9]. This advantage is novel and has recently been explored in studies across different AI systems.

Current research presents this variety of AI models developed to monitor and assist in different bilateral and unilateral^[8,10] IHR techniques, including transabdominal preperitoneal (TAPP), total extraperitoneal (TEP), and robotic inguinal hernia repairs (RIHR)^[8-10]. These AI applications optimize procedural efficiency and accuracy, significantly contributing to better surgical outcomes. Furthermore, the concurrent advancement of both laparoscopic and robotic techniques has been instrumental in advancing AI integration in hernia repair.

Due to the nascent nature of AI use in IHR, there is limited literature that compiles and aggregates the benefits and disadvantages of this practice. To the best of our knowledge, this review is the first systematic evaluation to compile and analyze the real-time intraoperative use of AI models in IHR. While recent discoveries show promising results, ongoing data collection and findings indicate that this area of practice and research is continuously evolving.

The purpose of this systematic review was to evaluate the latest use of AI models in the real-time intraoperative phase of IHR for their accuracy, applicability, and potential influence on clinical outcomes and surgical training. Additionally, the number of surgical phases recognized by each AI model is considered as a measure of phase-specific accuracy. This review provides a comprehensive and novel analysis of the recent performance and transformative potential of video-trained surgical phase-recognition AI systems, assessing their ability to enhance surgical outcomes and redefine traditional practices.

MATERIALS AND METHODS

RESULTS

Study selection

A comprehensive search was conducted across six databases, yielding 903 records in the initial literature search. After removing 169 duplicates, 734 records remained for screening. Based on pre-determined inclusion and exclusion criteria, we conducted a title and abstract review, selecting six studies for full-text analysis. Of these, one study was excluded because it utilized images for surgical phase recognition^[13]. After a thorough full-text review, data were extracted from five articles, all published in English. Additionally, a citation search was performed to ensure comprehensive coverage of relevant literature, identifying five additional references within the scope of our study. Upon reviewing these, one was excluded as it was a review article^[14], and three others were excluded for employing image-based surgical phase recognition methods^[15-17]. Consequently, one additional study was included in our analysis^[7]. The detailed reasons for exclusion and the study selection process are illustrated in the PRISMA flowchart [Figure 1], created using the online tool developed by Haddaway et al.^[18,19].

Figure 1. Search outputs based on PRISMA guidelines^[19]. PRISMA: Preferred Reporting Items for Systematic Review and Meta-Analysis.

DISCUSSION

In this systematic review, we evaluated studies utilizing AI and machine learning (ML) techniques to train models for real-time surgical phase recognition during minimally invasive IHR procedures. To our knowledge, this is the first systematic review to focus on this critical intersection of AI and surgical workflow, underscoring its potential to transform the future of surgery. AI holds immense promise for healthcare, offering opportunities to reduce labor demands and improve the quality of clinical analysis and decision making^[22]. Over the past few years, the role of artificial intelligence in hernia surgery has been examined from broad perspectives. For instance, Lima et al. (2024)^[23] performed a qualitative systematic review of 13 peer-reviewed studies published between 2020 and 2023. Their analysis suggests that machine learning platforms hold promise as tools for predicting surgical outcomes and identifying risk factors to mitigate postsurgical complications.

The integration of AI into healthcare is already reshaping patient care across multiple domains. For instance, AI-based systems like Babylon Health are enhancing both in-person and virtual consultations by providing timely and accurate assistance^[24]. Similarly, AI-powered tools such as Sense.ly, which combines a virtual assistant with an intuitive, human-like interface, are advancing medication management and health monitoring^[25] in AI-driven diagnostics. A particularly notable example of AI-driven diagnostics is Caption Health’s FDA-approved software, which enables healthcare providers to perform cardiac ultrasound imaging without extensive sonographic expertise. Acting as a “co-pilot,” this AI software offers real-time guidance, instructing users on transducer manipulation and ensuring the acquisition of diagnostic-quality images^[26].

In surgical practice, AI-driven phase recognition has emerged as a critical tool for workflow analysis and the evaluation of technical performance. Automation of operative phase segmentation and classification has the potential to train residents, improve skills, reduce iatrogenic complications, improve efficiency, and minimize costs. While various surgical specialties are exploring AI applications, laparoscopic procedures have been the primary focus of most studies. For example, significant advancements have been reported for surgeries such as laparoscopic sleeve gastrectomy^[27,28], sigmoidectomy^[29], myotomy^[30], and cholecyst-ectomy^[31].

AI functions primarily through ML, a paradigm that allows computers to learn and adapt to complex tasks without being explicitly programmed^[32]. Traditional algorithms, which rely on predefined rules, often fall short when tackling intricate problems in healthcare where variability and complexity abound. ML addresses this limitation by enabling computers to analyze large volumes of high-quality annotated data, iteratively improving their performance. Rather than relying on rigid rules to identify surgical phases, AI models are trained using vast annotated datasets, allowing algorithms to discern patterns and optimize phase recognition. Advanced techniques such as deep learning and neural networks further enhance AI’s capabilities by autonomously processing unstructured data, such as surgical videos, and extracting temporal and spatial dependencies to predict surgical workflows.

The Cholec80 dataset has been widely recognized as a benchmark for surgical phase recognition, providing a foundation for the development of deep learning models. Early attempts in 2016, such as EndoNet, achieved 75.2% accuracy by leveraging convolutional neural networks (CNN) for feature extraction combined with support vector machines and hidden Markov models for temporal phase inference^[9]. Subsequent advancements incorporated temporal dependencies, as seen with models like SV-RCNet, which utilized long short-term memory (LSTM) networks to achieve 81.6% accuracy in 2018^[33] and TeCNO (88.6%, 2020) using 3D convolutions^[34].

The introduction of attention mechanisms further enhanced the performance of deep learning models, as demonstrated by TMRNet (90.1%, 2021)^[35] and Trans-SVNet (90.3%, 2021)^[36]. Most recently, in 2023, a model incorporating 3D convolutions with positional encodings achieved an accuracy of 92.3%, representing the highest reported performance on the Cholec80 dataset to date^[37].

Building on these advancements, the current review highlights the expanding applicability of AI models across various surgical procedures, particularly in laparoscopic and robotic hernia repairs. In the study by Takeuchi et al. (2022), the TeCNO + HMM model, which integrates temporal convolutional networks and hidden Markov models, demonstrated accuracies of 88.81% for unilateral and 85.82% for bilateral TAPP repairs^[8]. These findings align closely with the performance of earlier attention-based models such as TMRNet and Trans-SVNet, which achieved accuracies of 90.1% and 90.3%, respectively, in cholecystectomy workflows^[35,36]. Similarly, Takeuchi et al. (2023) applied the transformer-based DETR model to analyze 211 TAPP videos, achieving an accuracy of 77.1%, with an F1 score of 75.4% and precision of 51.2%^[20]. While these results underscore the trade-off between precision and scalability, they also highlight the adaptability of transformer models in surgical phase recognition.

The study by Zygomalas et al. (2024) utilized the YOLOv8 deep learning model to analyze laparoscopic TAPP repair procedures, reporting an accuracy of 87.3% and an F1 score of 82%^[7]. However, unlike the TeCNO and ResNet-50 models, phase segmentation was not explicitly addressed, underscoring ongoing variability in study endpoints. Notably, the study by Choksi et al. (2023) implemented the ResNet-50 model in RIHR, reporting an accuracy of 68.7% and identifying seven procedural phases^[21]. This study also provided a critical latency period of 250 milliseconds, a key consideration for real-time applications, especially when paired with edge computing devices like the Nvidia Jetson Xavier™.

In contrast, Zang et al. (2023) analyzed RALIHR using a CV/DL (Computer Vision/Deep Learning) model, achieving an accuracy of 85% with an F1 score of 78.51%^[9]. This study also utilized Nvidia A100 as its edge computing device, further highlighting the role of hardware advancements in improving model performance. Finally, the study by Ortenzi et al. (2023) stands out for its scale, analyzing 619 TEP repairs across multiple centers and involving 85 surgeons^[10]. By employing the Surgical Intelligence Platform incorporating Video Transformer Networks (VTN) and LSTM networks, this study achieved a mean accuracy of 88.8%, with 89.59% for unilateral cases and 88.45% for bilateral cases. These findings reaffirm the significance of large-scale datasets and multi-institutional collaborations in enhancing AI model generalizability and minimizing bias.

Despite promising results, this systematic review has several limitations. First, the included studies exhibited considerable heterogeneity in terms of AI model architectures (e.g., CNN-based models like ResNet-50, transformer-based models such as DETR and VTN), dataset sizes (ranging from 25 to 619 videos), surgical approaches (TAPP, TEP, RIHR), and annotation protocols (manual vs. semi-automated tools such as CVAT and SuperAnnotate). This heterogeneity extended to evaluation metrics, with accuracy ranging from 74% to over 87%, and F1 scores from 65% to 82%, with some studies lacking complete reporting of precision and recall. Such variability not only precluded meta-analysis but also limited the comparability and generalizability of model performance across surgical contexts.

Second, internal validation was lacking or inconsistently reported in several studies. The absence of rigorous methods such as k-fold cross-validation or independent test sets raises concerns regarding overfitting and diminishes the robustness of the reported outcomes. Third, most studies were observational and retrospective in design, introducing selection bias and limiting the applicability of findings to real-time, prospective surgical workflows. Fourth, surgeons’ experience, both in performing the procedures and in annotating surgical videos, was either underreported or not described in most studies. This omission may influence the quality of annotations and affect model training and reliability, especially for complex or skill-dependent phases. Finally, manual annotation is labor-intensive and time-consuming, and inconsistencies in labeling standards, phase definitions, and inter-annotator agreement can negatively impact model reproducibility. Lastly, institutional variability in operative techniques, equipment, and surgical team experience further limits the external validity and clinical scalability of the current AI models reviewed.

Future research should prioritize the development of standardized, publicly available, multi-institutional datasets with consistent surgical phase definitions and annotation protocols. This would enable robust benchmarking of AI models across institutions and procedural techniques. Establishing open repositories for annotated surgical videos, including TAPP, TEP, and RIHR procedures, could facilitate reproducibility and fairness in model comparison. In addition, integrating quantitative measures of heterogeneity, such as I² statistics, τ² values, and the coefficient of variation of reported accuracies into future reviews could help identify underlying sources of inconsistency and guide model selection.

Advanced training strategies such as federated learning and self-supervised approaches may reduce the dependence on manually labeled data while ensuring data privacy across institutions. Further research should focus on the development of real-time AI systems with low-latency performance using edge computing hardware such as Nvidia Jetson or A100 GPUs. These systems should be validated prospectively in real-world, multicenter clinical settings and evaluated not only for model accuracy but for their impact on surgical workflow, efficiency, safety, and outcomes.

Equally important is the pursuit of explainable AI to foster trust and transparency. Future models should incorporate interpretable outputs, including heatmaps, attention weights, and confidence scores, to assist surgical decision making and training. Regulatory pathways, including alignment with FDA and international guidelines on software as a medical device, must be actively considered during development and deployment. Interdisciplinary collaboration among surgeons, engineers, regulatory authorities, and ethicists will be crucial in ensuring safe, effective, and equitable integration of AI into operative practice.

From our perspective, this systematic review represents a foundational step toward shaping the future of AI in surgical education and operative care. We envision that real-time AI-assisted surgical phase recognition will become an essential tool in residency training, competency evaluation, and intraoperative decision support. For early-career robotic surgeons and general surgery trainees, such systems could offer objective, adaptive feedback that enhances skill acquisition. The future of surgical AI lies not only in advancing algorithmic performance but also in establishing the human-AI partnership at the core of modern surgical practice. With early, coordinated investment in algorithm development, dataset curation, and prospective validation, we believe this vision can be realized.

In conclusion, This systematic review underscores the transformative potential of AI and ML in advancing real-time surgical phase recognition during minimally invasive IHR. While significant progress has been achieved across laparoscopic and robotic procedures, the application of AI in IHR remains in its early stages. The included studies demonstrate encouraging performance metrics, with models like TeCNO, ResNet-50, and Video Transformer Networks achieving promising accuracies. However, challenges such as heterogeneity in model architectures, reliance on labor-intensive annotated datasets, and variability in surgical techniques must be addressed to enhance the clinical utility of AI. Future research should focus on developing standardized benchmarks, conducting prospective multicenter validations, and leveraging hardware advancements to enable latency-free, real-time applications. Collaborative efforts among surgeons, AI developers, and policymakers will be pivotal to overcoming existing limitations, ensuring ethical implementation, and integrating AI-driven systems into surgical workflows. By doing so, AI has the potential to revolutionize surgical training, improve technical performance, and ultimately enhance patient outcomes in IHR and beyond.