Artificial intelligence in endoscopic ultrasound-guided biliary drainage: Emerging applications, technical challenges, and future directions

Abstract

Endoscopic ultrasound-guided biliary drainage (EUS-BD) has evolved from a rescue procedure following failed endoscopic retrograde cholangiopancreatography (ERCP) into an important therapeutic option for select patients with biliary obstruction. However, EUS-BD remains technically demanding and highly operator-dependent, requiring accurate anatomical recognition, safe route selection, stable access, and precise stent deployment. Artificial intelligence (AI) has the potential to improve the safety, reproducibility, and efficiency of EUS-BD by supporting procedural decision making. This review discusses emerging AI applications in EUS-BD, including the prediction of ERCP failure, pre-procedural imaging assessment, drainage route selection, real-time EUS image interpretation, bile duct recognition, vessel avoidance, needle trajectory planning, device tracking, stent deployment support, adverse event prediction, and post-procedural follow-up. AI can further assist in training and quality assurance through anatomical labeling, procedural feedback, and objective performance assessment. However, direct evidence for AI-guided EUS-BD remains limited, and most AI-EUS studies have focused on diagnostic rather than therapeutic applications. The major barriers include limited datasets, a lack of expert-labeled procedural videos, inadequate external validation, challenges in real-time integration, limited model explainability, and ethical and regulatory uncertainties. AI-assisted EUS-BD should therefore be considered as investigational rather than standard practice. Future progress requires multicenter validation studies and clinically meaningful evidence to demonstrate improved procedural performance and patient outcomes.

Key Words: Artificial intelligence; Endoscopic ultrasound; Endoscopic ultrasound-guided biliary drainage; Therapeutic endoscopic ultrasound; Biliary obstruction

Core Tip: Artificial intelligence (AI) could become an important add-on to endoscopic ultrasound (EUS)-guided biliary drainage by supporting patient selection, pre-procedural planning, anatomical recognition, real-time procedural guidance, and adverse-event prediction. Although direct clinical evidence remains limited, AI has the potential to enhance procedural precision, reduce operator dependence, and improve training during this technically demanding procedure. Future adoption will necessitate multicenter datasets, prospective validation, explainable models, and seamless integration into the therapeutic EUS workflow.

INTRODUCTION

Malignant biliary obstruction is a common and clinically significant consequence of pancreatic cancer, cholangiocarcinoma, ampullary cancer, and other periampullary malignancies. Effective biliary drainage is essential for relieving jaundice, preventing or treating cholangitis, improving hepatic function, and facilitating systemic anticancer therapy. Endoscopic retrograde cholangiopancreatography (ERCP) remains the standard first-line approach for biliary decompression; however, it may fail because of duodenal obstruction, tumor infiltration of the papilla, surgically altered anatomy, periampullary diverticula, or inability to achieve selective biliary cannulation[1,2].

Endoscopic ultrasound-guided biliary drainage (EUS-BD) has emerged as an important alternative to percutaneous transhepatic biliary drainage (PTBD) and surgical bypass after failed ERCP. Since its first description in 2001, EUS-BD has evolved from an experimental rescue procedure to an established therapeutic option in expert centers[3].

Current EUS-BD techniques include endoscopic ultrasound (EUS)-guided choledochoduodenostomy, hepaticogastrostomy, antegrade stenting, and the rendezvous approach. These techniques enable internal BD without traversing the papilla and may reduce some ERCP-related adverse events, particularly post-ERCP pancreatitis and tumor-related stent dysfunction[2,4].

Recent comparative studies have generated interest in EUS-BD not only as a rescue therapy but also as a primary drainage strategy in eligible patients with malignant distal biliary obstruction. Meta-analytic reviews of randomized controlled trials have indicated that EUS-BD and ERCP-guided drainage have comparable technical and clinical success rates, whereas EUS-BD may be associated with lower re-intervention rates, a reduced risk of post-procedural pancreatitis and tumor ingrowth or overgrowth, and shorter hospital stays[4].

However, EUS-BD remains technically demanding and highly operator dependent. Accurate recognition of the biliary and vascular anatomy, appropriate target-duct selection, controlled scope positioning, safe needle-trajectory planning, appropriate device selection, and accurate stent deployment are pivotal for successful performance[4]. Artificial intelligence (AI), particularly through computer vision, deep learning techniques, has expanded rapidly in gastrointestinal endoscopy, enabling applications such as image segmentation, lesion classification, and real-time quality control. In EUS, AI has been studied predominantly for diagnostic applications, including the assessment of pancreatic masses, differentiation of pancreatic cancer from inflammatory lesions, stratification of the risk of pancreatic cysts, and localization of anatomical landmarks[5,6].

In this regard, deep learning systems for bile duct annotation and station recognition during linear EUS[7] represent an emerging advancement that demonstrates the feasibility of AI-assisted biliary structure recognition and EUS scanning augmentation. These advances indicate that AI may one day aid in diagnostic EUS and technically complex clinical processes requiring high spatial awareness and procedural precision.

Currently, the application of AI to EUS-BD remains in its early stages. AI is not currently routinely used to guide EUS-BD in clinical practice and direct evidence remains limited. To avoid overstating the current evidence, AI applications in this field can be classified into three categories. Validated AI applications in EUS are mainly diagnostic and include image interpretation, lesion characterization, and anatomical recognition. Proof-of-concept applications include bile duct annotation, station recognition, and early feasibility work in pancreaticobiliary imaging. In contrast, AI-assisted drainage-route selection, needle trajectory planning, vessel avoidance, device tracking, stent deployment support, and real-time therapeutic guidance during EUS-BD remain speculative future developments and have not yet been validated for routine clinical practice. Most published AI studies in EUS have focused on diagnostic applications rather than on real-time guidance of therapeutic EUS procedures, and only limited data are available specifically for EUS-BD.

A clear distinction should also be made between conventional machine-learning and deep-learning approaches and large language models (LLMs). Most AI applications discussed in the technical EUS-BD workflow, including image interpretation, anatomical recognition, drainage-route selection, vessel avoidance, needle trajectory planning, device tracking, and procedural guidance, are primarily based on computer vision, image segmentation, classification algorithms, and other machine-learning or deep-learning methods rather than LLMs. However, LLMs may have a different and complementary future role in EUS-BD, including automated report generation, structured procedural documentation, clinical decision support, literature synthesis, training assistance, patient-specific summarization, and integration of multimodal clinical information. However, these LLM-based applications also remain investigational and require validation, safeguards against inaccurate outputs, clinician oversight, and integration into existing endoscopy workflows before clinical implementation.

Nevertheless, EUS-BD is an interesting target for AI-assisted therapeutic endoscopy because many procedural steps are image- and anatomy-dependent and risk-sensitive. Potential applications could include predicting ERCP failure; assessing the choice between ERCP, EUS-BD, and percutaneous drainage; assessing pre-procedural cross-sectional imaging; recognizing bile ducts and adjacent vessels in real time; selecting route; guiding needle trajectory; supporting stent deployment; and predicting adverse events, including bleeding, bile leak, stent migration, cholangitis, and re-intervention. These potential applications should currently be regarded as conceptual or investigational, rather than as tools approaching routine clinical implementation. In this review, we describe the nascent role of AI in EUS-BD, which may find applications in patient selection, procedural planning, image interpretation, real-time guidance, safety prediction, and post-procedural follow-up. We also highlight the technical, ethical, regulatory, and implementation challenges that must be addressed before AI-assisted EUS-BD becomes a clinical reality.

CURRENT ROLE OF EUS-BD

EUS-BD was originally developed as a rescue intervention after failed ERCP but has since evolved into a central component of therapeutic EUS. It continues to play an integral role in the treatment of biliary obstruction when effective transpapillary drainage is unsuccessful or infeasible, particularly in malignant distal biliary obstruction. ERCP failure may result from tumor infiltration of the papilla, duodenal obstruction, surgically altered anatomy, or difficult biliary cannulation. In this respect, EUS-BD offers internal biliary decompression under ultrasound guidance without the use of external drainage catheters, thereby mitigating PTBD morbidity[8,9].

The main EUS-BD methods include endoscopic ultrasonography-guided choledochoduodenostomy, hepaticogastrostomy, antegrade stenting, and rendezvous drainage. EUS-guided choledochoduodenostomy is commonly performed from the duodenal bulb to the extrahepatic bile duct, and is most commonly used for malignant distal biliary obstruction. EUS-guided hepaticogastrostomy is performed by puncturing the dilated left intrahepatic duct from the stomach, followed by the creation of a hepaticogastric fistula and subsequent stent placement. This method may be particularly helpful in patients with duodenal obstruction or surgically altered anatomy. Antegrade stenting and rendezvous techniques require EUS-guided biliary access, followed by the passage of a guidewire across the papilla or stricture to enable internal drainage while preserving or restoring transpapillary access[10,11].

In contrast to PTBD, EUS-BD offers several practical and patient-focused benefits. Although PTBD is effective, external drainage catheters pose a risk of quality-of-life impairment and can result in catheter dislodgement, bile leakage, skin site infection, bleeding, and recurrent interventions. Meta-analyses comparing EUS-BD and PTBD following ERCP failure suggest that EUS-BD yields similar technical outcomes with lower overall complication and re-intervention rates in suitably selected patients[9,12].

In addition to its role as a rescue therapy, the prospective clinical applications of EUS-BD have expanded. Its potential as a primary drainage alternative for malignant distal biliary obstruction has been evaluated in comparison with ERCP. Potential benefits include the avoidance of papillary manipulation, decreased risk of post-ERCP pancreatitis, internal drainage from the tumor-infiltrating papilla, and lower rates of tumor ingrowth or overgrowth. Recent comparative studies have indicated that EUS-BD may demonstrate clinical efficacy comparable to that of ERCP, with lower rates of re-intervention and post-procedural pancreatitis in selected patients. However, primary EUS-BD has not achieved universal adoption because of the need for advanced expertise, dedicated devices, precise patient selection, and reliable management strategies for adverse events[13].

In particular, performing EUS-BD is a technically demanding procedure. Successful drainage requires correct interpretation of cross-sectional imaging and real-time EUS anatomy, identification of vascular structures, selection of a target puncture location, maintenance of a stable endoscope position, controlled tract dilation, and accurate stent insertion. Potential adverse events include bile leak, bleeding, pneumoperitoneum, perforation, cholangitis, peritonitis, stent migration, stent dysfunction, and re-intervention. These issues highlight the need for EUS-BD to be performed only in specialized centers, with continued emphasis on training, procedural standardization, and pre-procedural planning[8,13].

WHY AI MAY BE USEFUL IN EUS-BD

EUS-BD is a multifactorial treatment, in which success depends on several interdependent decisions rather than on a single technical maneuver. Endoscopists must determine whether EUS-BD is indicated, assess the safest route for drainage, identify the target bile duct, avoid intervening vessels, choose an optimal puncture trajectory, maintain an appropriate endoscopic position, and accurately place the stent. Patient anatomy, tumor localization, ductal dilatation, duodenal accessibility, ascites, coagulopathy, and local expertise play roles in each of these steps. The heterogeneity of the human anatomy and procedural complexity in EUS-BD renders it a good candidate for AI-assisted decision support and machine learning image-guided procedural optimization[14].

Image interpretation and anatomical recognition are likely among the first use cases of AI. EUS-BD entails accurate localization of the extrahepatic or intrahepatic bile ducts, portal venous structures, hepatic vessels, tumor infiltration, and anatomical relations between the gastrointestinal wall and target duct. Deep-learning systems have been applied in EUS to aid bile duct annotation and standard station identification in linear EUS tests. Although such systems are not specifically designed for therapeutic drainage, they demonstrate that AI can recognize biliary structures on EUS images and may eventually assist with real-time anatomical labeling during EUS-BD[15]. AI might also assist in the selection of drainage routes. The choice between EUS-guided choledochoduodenostomy, hepaticogastrostomy, antegrade stenting, or rendezvous drainage currently depends significantly on operator preference, ductal anatomy, site of obstruction, duodenal patency, and expected technical feasibility. AI systems that combine endoscopic findings with EUS images and computed tomography (CT) and/or magnetic resonance imaging (MRI) data, as well as laboratory parameters and prior ERCP data, may inform the optimal drainage route for individual patients. Such systems may be particularly useful in borderline clinical scenarios, such as partial duodenal obstruction, surgically altered anatomy, limited ductal dilatation, or elevated risk of adverse events[16].

Another potential application is risk prediction. The adverse events associated with EUS-BD include bile leakage, bleeding, cholangitis, perforation, pneumoperitoneum, stent migration, and recurrent biliary obstruction. Procedural risks, such as intervening vessels or unstable scope positioning, are often evident in real-time whereas others are influenced by pre-procedural factors, such as ascites, coagulopathy, tumor burden, duct diameter, and patient performance status. AI-based prediction models may be utilized to stratify procedural risk, identify patients who may benefit from alternative drainage strategies, and guide the implementation of additional safety measures, such as antibiotic optimization, correction of coagulopathy, or selection of alternative stents[17].

AI may also assist in training and procedural standardization. EUS-BD has a steep learning curve and remains concentrated in high-volume expert centers. Variability in techniques, device selection, and complication management limit the reproducibility of results across institutions. AI-assisted feedback systems can assist trainees identify anatomical landmarks, confirm scanning stations, highlight unsafe puncture windows, and provide objective performance metrics. Additionally, AI systems may eventually be integrated with simulation-based platforms to provide structured EUS-BD training before independent clinical practice[18].

Currently, AI in EUS-BD must be considered an emerging concept, rather than an established clinical tool. Most AI applications in EUS have focused on diagnostic tasks, predominantly the detection and classification of pancreatic lesions, rather than therapeutic guidance. Direct evidence supporting the use of AI in EUS-BD is limited and no AI platform is currently available for real-time BD. Thus, AI’s current applications in EUS-BD primarily focus on enhancing procedural planning, anatomical recognition, risk prediction, training, and reproducibility. Potential roles in drainage-route selection, unsafe-window detection, puncture-trajectory planning, device tracking, and stent-deployment support should presently be viewed as investigational directions rather than near-term clinical applications. Prospective studies are required to assess whether this technology translates to quantifiable clinical improvements[16,18].

AI FOR PATIENT SELECTION AND PRE-PROCEDURAL PLANNING

For EUS-BD to succeed, appropriate patient selection is key. Currently, in EUS-BD evaluation, management decisions depend on several aspects, including clinical status, indication for biliary decompression, likelihood of ERCP success, anatomical feasibility, local expertise, and availability of alternative drainage strategies. AI-based predictive models are among the approaches that may assist in this process by integrating clinical, biochemical, endoscopic, and imaging factors to estimate the risk of ERCP failure, clinical appropriateness of EUS-BD, and likelihood of requiring rescue drainage[19].

One important clinical application of such models is the pre-procedural prediction of ERCP difficulty or failure. Malignant duodenal obstruction, papillary tumor infiltration, surgically altered anatomy, proximal biliary obstruction, periampullary diverticula, and previous unsuccessful cannulation can lower the chances of successful transpapillary drainage. Machine learning models trained on large ERCP datasets could potentially identify patients for whom early EUS-BD might be preferred over repeat ERCP attempts, thus reducing procedure time, minimizing papillary trauma, and shortening the interval to effective biliary decompression[19,20].

AI may also support the selection of the most appropriate drainage approach once EUS-BD is under consideration. EUS-guided choledochoduodenostomy, hepaticogastrostomy, antegrade stenting, or rendezvous drainage can be performed depending on the severity of the obstruction, degree of ductal dilatation, duodenal access, presence of gastric outlet obstruction or ascites, vascular anatomy, and operator expertise. With pre-procedural CT, MRI/magnetic resonance cholangiopancreatography (MRCP), prior ERCP findings, and EUS images integrated into AI systems, future models may help assess the most appropriate approach to achieving optimal access in individual patients, potentially identify the safest access routes, and predict clinical success using these techniques. However, this application remains largely conceptual, with no direct validation in EUS-BD-specific populations[20,21].

Pre-procedural planning is closely associated with cross-sectional imaging findings. CT and MRI/MRCP are used to identify the degree and extent of biliary obstruction, distribution of intrahepatic and extrahepatic ductal dilatation, tumor burden, vascular involvement, ascites, and surgically altered anatomy. Developing deep learning approaches for biliary tree segmentation and 3D reconstruction may ultimately enable automated mapping of the biliary anatomy before EUS-BD. These tools may enable endoscopists to visualize the target duct, avoid intervening vessels or tumors, and plan puncture trajectories more accurately[22].

AI can also assist in pre-procedural risk stratification. Ascites, coagulopathy, poor performance status, non-dilated ducts, hilar obstruction, extensive tumor infiltration, or unstable anatomy may increase the risk of bile leakage, bleeding, stent dysfunction, or clinical failure in patients. Predictive models can classify patients into low, intermediate, and high-risk groups and help clinicians determine whether to proceed with EUS-BD, optimize modifiable risk factors, involve early interventional radiology or surgery, or consider alternative drainage approaches, such as PTBD[23].

Despite these potential benefits, the application of AI in patient selection for EUS-BD remains largely conceptual. Most AI studies in pancreaticobiliary endoscopy have focused on diagnosis rather than therapeutic decision making; only a few predictive models have been prospectively validated in EUS-BD populations. Therefore, future research must prioritize multicenter datasets, transparent model development, external validation, and clinically meaningful outcomes, such as technical success, clinical success, adverse events, re-intervention, length of hospital stay, and time to oncological therapy[19,21].

AI FOR IMAGE INTERPRETATION AND REAL-TIME PROCEDURAL GUIDANCE

As the ability to interpret and process real-time images is the foundation of EUS-BD, anatomical recognition directly influences duct puncture, guidewire manipulation, tract dilation, and stent deployment. AI may improve the detection of the bile duct, vascular structures, tumors, gastrointestinal wall, and other intervening anatomies in technically challenging procedural steps, thereby supporting safer and more efficient decision-making. However, current evidence mainly supports AI-assisted anatomical recognition and image interpretation, whereas direct evidence for AI-guided therapeutic EUS-BD remains limited[24].

Bile duct annotation and station recognition are feasible using deep learning systems for linear EUS examinations. Although these models were developed for diagnostic EUS rather than therapeutic drainage, they provide an important foundation for future AI-assisted anatomical recognition in EUS-BD[25].

AI can also assist with puncture site and trajectory selection by combining EUS videos, Doppler imaging, CT or MRI data, and device position information. Potential applications, such as recognizing safe access windows, estimating the duct-to-wall distance, detecting intervening vessels, and proposing stable access trajectories, are still under investigation and have not yet been clinically validated for EUS-BD[26].

Future AI platforms could also include guidewires, dilators, and stent tracking using EUS or fluoroscopy, with automated alerts for unstable alignment or suboptimal stent deployment. Such capabilities may be significant during hepaticogastrostomy and antegrade procedures in which stent maldeployment can result in serious complications. Currently, however, these applications remain speculative and are not ready for routine clinical implementation[27].

The establishment of a procedural map integrated with the bile duct, vascular structures, tumor, gastrointestinal wall, and planned access routes using AI-based fusion of EUS, fluoroscopy, and cross-sectional imaging may also be beneficial. Such systems may help reduce the need for mental three-dimensional reconstruction using multiple 2D imaging modalities, particularly during more complex interventions[28].

Nonetheless, real-time AI guidance is technically challenging despite its potential benefits. EUS images are typically operator-dependent and may be affected by artifacts, respiratory motion, acoustic shadowing, and substantial anatomical variability. From an engineering perspective, real-time models require frame quality control, motion-robust segmentation, temporal smoothing across video frames, and training on variable probe positions, gastrointestinal peristalsis, and unstable probe–tissue contact to maintain reliable anatomical recognition during live EUS-BD. Temporal sequence-modeling approaches used in surgical video analysis, including 3D convolutions, optical flow methods, and transformer-based architectures, may help AI systems leverage continuous frame information rather than isolated static images, thereby improving structure tracking and reducing frame-to-frame uncertainty. Therefore, developing robust AI systems requires expert-annotated procedural videos, standardized procedural phase labeling, and prospective validation across multiple platforms, devices, centers, and operators[29].

Hence, AI should be considered as an adjunct to expert operator judgment and not a substitute for procedural expertise. Short- to medium-term use will likely include anatomical labeling, bile duct segmentation, vessel warning systems, and training feedback, whereas widespread clinical adoption will require evidence demonstrating improvements in technical success, procedural safety, procedural time, re-intervention rates, and patient outcomes[24,29]. The potential applications of AI across the EUS-BD workflow are summarized in Table 1, and the proposed AI-assisted procedural workflow is illustrated in Figure 1.

Figure 1 Proposed workflow for artificial intelligence-assisted endoscopic ultrasound-guided biliary drainage, showing potential roles across patient selection, imaging review, route selection, real-time image interpretation, needle trajectory and stent deployment support, adverse-event prediction, and follow-up planning. AI: Artificial intelligence; EUS: Endoscopic ultrasound; EUS-BD: Endoscopic ultrasound-guided biliary drainage; CT: Computed tomography; MRI-MRCP: Magnetic resonance imaging-magnetic resonance cholangiopancreatography; CDS: Choledochoduodenostomy HGS: Hepaticogastrostomy.

Table 1 Potential applications of artificial intelligence in endoscopic ultrasound-guided biliary drainage.

Stage of care	AI application	Data input	Potential clinical benefit	Current limitation	Ref.
Patient selection	Prediction of ERCP failure and suitability for EUS-BD	Clinical history, laboratory values, prior ERCP findings, CT/MRI findings	Earlier selection of EUS-BD and avoidance of repeated failed ERCP attempts	Requires large, validated therapeutic EUS datasets	[19,20]
Pre-procedural planning	Selection of drainage route, including CDS, HGS, antegrade stenting, or rendezvous drainage	CT, MRI/MRCP, EUS images, obstruction level, ductal dilatation, duodenal patency	Individualized selection of the safest and most feasible drainage strategy	Limited direct evidence in EUS-BD-specific populations	[20-23]
Anatomical recognition	Identification of bile duct, vessels, tumor, and gastrointestinal wall	Real-time EUS images, EUS video, Doppler EUS, cross-sectional imaging	Improved target recognition and safer access planning	Most current AI-EUS studies remain diagnostic or proof-of-concept	[24,25]
Route and trajectory guidance	Suggestion of puncture window and needle trajectory	EUS video, Doppler signal, duct-wall distance, device position	Reduced risk of bleeding, bile leak, and failed access	Requires real-time integration into EUS processors and devices	[26,27]
Device and stent deployment support	Recognition of guidewire, dilator, and stent position during deployment	EUS images, fluoroscopy, procedural video, device-tracking data	More precise stent placement and reduced maldeployment risk	Not yet established in routine clinical practice	[26,27]
Safety optimization	Prediction of bile leak, bleeding, cholangitis, stent migration, stent dysfunction, and reintervention	Clinical variables, coagulation profile, ascites, duct diameter, imaging and procedural data	Better consent, risk mitigation, and post-procedure monitoring	Needs prospective validation using clinically meaningful endpoints	[17,23]
Training and quality assurance	Landmark recognition, procedural phase detection, and objective performance feedback	EUS recordings, procedure videos, annotated landmarks, procedural metrics	Standardized training, competency assessment, and reduced operator variability	Requires expert-labelled datasets and standardized competency metrics	[18,24,25]
Post-procedural follow-up	Prediction of clinical failure, stent dysfunction, or need for reintervention	Symptoms, bilirubin trend, liver enzymes, imaging, prior procedural details	Earlier detection of treatment failure and timely reintervention	Minimal evidence currently available specifically for EUS-BD	[17,23,27]

AI: Artificial intelligence; BD: Biliary drainage; CDS: Choledochoduodenostomy; ERCP: Endoscopic retrograde cholangiopancreatography; EUS: Endoscopic ultrasound; HGS: Hepaticogastrostomy; MRI: Magnetic resonance imaging; MRCP: Magnetic resonance cholangiopancreatography.

SAFETY, LIMITATIONS, AND IMPLEMENTATION CHALLENGES

However, safety remains an important issue for EUS-BD. The procedure requires transmural duct puncture, guidewire manipulation, tract dilation, and stent deployment. Potential adverse events include bile leak, bleeding, perforation, cholangitis, stent migration, recurrent biliary obstruction, and the need for re-intervention. AI-based risk-prediction models may gradually offer new methods for identifying high-risk patients and supporting safer procedural planning[30].

AI has the potential to enhance procedural safety by incorporating clinical variables, laboratory data, CT or MRI findings, EUS imaging, Doppler assessment, and fluoroscopic information to predict technical failures or adverse events. The real-time recognition of bile ducts, vascular structures, device position, and unsafe puncture windows can theoretically minimize the risk of bleeding, bile leakage, stent maldeployment, and failed access[31].

However, the evidence remains limited, as most AI-EUS studies have examined diagnostic use rather than therapeutic guidance, and few datasets include full EUS-BD procedural videos, failed procedures, device-manipulation sequences, near-miss events, or long-term clinical outcomes. Findings from diagnostic AI-EUS studies cannot be assumed to translate directly into safe real-time guidance for therapeutic EUS-BD. The establishment of clinically reliable systems requires multicenter datasets, expert-annotated procedural videos, and rigorous external validation before routine deployment[32].

Training, ethics, and integration into clinical workflow remain significant barriers. Although AI may assist in identifying anatomical milestones and providing procedural feedback, it cannot replace structured therapeutic EUS training, expert supervision, or operator judgment. Explainable models, well-defined accountability structures and safety mechanisms, data security protection, prospective validation studies, and the integration of real-time clinical workflows are necessary before AI-assisted EUS-BD can be integrated into routine clinical practice[33,34].

CONCLUSION

The future of AI-assisted EUS-BD relies on the development of large multicenter datasets that include clinical variables, cross-sectional imaging, EUS videos, fluoroscopy, procedural details, adverse events, and longitudinal follow-up outcomes. These datasets should include not only successful procedures but also failed attempts, near-miss events, and complications to ensure that AI systems are trained on real-world procedural variability rather than idealized cases[35]. Including failure cases is crucial because they teach models to recognize unsafe anatomy, unstable access, device misalignment, and early warning patterns that may precede technical failure or adverse events.

Cross-center generalizability will also require attention to data drift caused by differences in EUS platforms, probe frequencies, imaging settings, operator expertise, and patient populations; strategies, such as domain adaptation and federated learning may help improve model robustness while preserving data privacy.

Prospective validation is essential before clinical implementation. LLMs should also be considered as a separate AI category within future therapeutic EUS workflows. Unlike computer vision and deep learning tools that may support image interpretation, anatomical recognition, and procedural guidance, LLMs are more likely to assist with structured report generation, procedural documentation, clinical decision support, literature synthesis, training, and the integration of multimodal clinical information. However, the use of LLMs in EUS-BD remains investigational and requires robust evaluation, safeguards against inaccurate outputs, data-security protections, and clinician oversight. Future studies should determine whether AI-assisted EUS-BD improves clinically meaningful outcomes, including technical success, clinical success, procedure time, adverse events, re-intervention rates, length of hospital stay, and time to oncological therapy. Explainable AI models are also critical, particularly for high-risk therapeutic procedures, where clinicians need to understand the rationale underlying AI-generated recommendations[36].

The application of AI along with real-time EUS platforms, fluoroscopy, and cross-sectional imaging, may eventually assist in drainage route selection, anatomical labeling, vessel avoidance, needle trajectory planning, device tracking, and stent deployment. Simultaneously, AI-based simulation platforms and procedural feedback systems may enhance training and contribute to standardized competency assessments in therapeutic EUS[37]. Recent (2026) publications on LLM applications in gastroenterology and multimodal endoscopic image interpretation further highlight the rapid evolution of language-based AI and the need for careful validation before clinical implementation[38,39].

In conclusion, AI in EUS-BD remains an emerging field of research. Although direct clinical evidence remains limited, AI has the potential to support patient selection, procedural planning, image interpretation, safety prediction, training, and post-procedural follow-up. However, proposed AI applications should be interpreted according to their current level of evidence: Validated diagnostic EUS applications, proof-of-concept anatomical or imaging tools, and speculative future therapeutic guidance systems. Most of the currently available evidence is derived from diagnostic EUS rather than AI-guided therapeutic EUS. Therefore, applications, such as route selection, needle trajectory planning, device tracking, vessel avoidance, and stent deployment support should be regarded as future research goals rather than established or near-clinical tools. In addition, computer vision and machine-learning approaches should be distinguished from LLM-based tools, which are more likely to support documentation, decision support, training, and multimodal information integration rather than direct real-time procedural guidance. The future role of AI in EUS-BD will depend on the availability of high-quality datasets, prospective validation, explainable models, seamless workflow integration, and clear evidence that AI-assisted EUS-BD improves patient outcomes beyond those achieved in expert endoscopic practice.