Artificial intelligence for fluoroscopic cholangiogram interpretation during endoscopic retrograde cholangiopancreatography

Abstract

Endoscopic retrograde cholangiopancreatography (ERCP) is a dominant therapeutic approach for biliary and pancreatic ductal diseases, but it remains technically demanding, operator-dependent, and primarily fluoroscopy-guided. The cholangiogram produced during each ERCP is the central image on which intraprocedural decisions rest, but it is interpreted subjectively, with appreciable interobserver variability. Unlike luminal mucosa or cross-sectional imaging, fluoroscopic cholangiography has remained a relatively underexploited area for artificial intelligence (AI). This narrative review outlines the potential applications of AI, particularly deep learning, in interpreting fluoroscopic cholangiograms during ERCP. Reported applications include differentiating malignant from benign biliary strictures, detecting and quantifying common bile duct stones, automated scoring of stone-extraction difficulty, stent-length selection, real-time cannulation and papilla guidance, predicting post-ERCP pancreatitis, and reducing radiation dose. Although early evidence is promising across several of these domains, it is largely based on small, single-center, retrospective datasets, with limited external validation and no regulatory-approved systems. Because erroneous output may immediately influence an invasive decision, the most credible near-term role is assistive: Confidence-aware, physician-in-the-loop support evaluated prospectively for safety, workflow, generalizability, governance, equity, privacy, and patient-important outcomes.

Key Words: Artificial intelligence; Biliary tract; Cholangiography; Endoscopic retrograde cholangiopancreatography; Fluoroscopy; Patient safety

Core Tip: Artificial intelligence may help transform fluoroscopic cholangiogram interpretation during endoscopic retrograde cholangiopancreatography (ERCP) from subjective visual assessment to objective, reproducible decision support. Current applications include biliary stricture characterization, bile duct and stone segmentation, stone-extraction difficulty scoring, stent-length selection, papilla and cannula localization, post-ERCP pancreatitis prediction, and potential radiation-dose reduction. However, most evidence remains retrospective and early-stage. Future progress requires prospective validation, external testing, explainability, workflow integration, data governance, and multimodal physician-in-the-loop systems that combine fluoroscopy with endoscopic, endosonographic, and clinical data.

INTRODUCTION

Endoscopic retrograde cholangiopancreatography (ERCP) is a cornerstone therapeutic procedure for biliary and pancreatic ductal disease, but it remains one of the most technically challenging and operator-dependent procedures in gastrointestinal endoscopy[1]. In contrast to luminal endoscopy, which allows direct examination of the mucosa, ERCP relies largely on fluoroscopy, and the cholangiogram serves as the principal map by which the endoscopist localizes pathology, plans cannulation, and selects therapy[1,2]. The interpretive burden is high. Interpretive error may contribute to incomplete stone clearance, missed malignant obstruction, inappropriate device selection, or procedure-related harm, although its contribution has not been quantified[1,2]. As case complexity increases and access to advanced expertise remains uneven, interest is increasing in tools that could support more objective and reproducible interpretation of cholangiographic images.

Despite its central role, the cholangiogram remains subjective, experience-dependent, and subject to interobserver variability. Indeterminate biliary strictures remain difficult to characterize confidently on imaging alone, and assessment of stone size, number, and ductal configuration, which are key factors influencing extraction difficulty, remains largely subjective and experience-dependent[2,3]. These limitations are not failures of the modality itself, but rather reflect the difficulty of extracting quantitative information from a low-contrast, two-dimensional projection in real time.

Artificial intelligence (AI), particularly deep learning, has been increasingly adopted in luminal endoscopy, with computer-aided detection systems for colorectal neoplasia having entered clinical practice in some settings[3]. Nevertheless, in pancreaticobiliary endoscopy, AI remains at a nascent stage, and the literature to date has focused mainly on endoscopic ultrasound, digital cholangioscopy, and white-light images of the papilla rather than on the fluoroscopic image itself[1,3]. Thus, the cholangiogram, which is generated during ERCP and may be stored as still images or video, remains largely underexploited.

This is beginning to change. Initial studies have shown that deep learning models can differentiate malignant from benign biliary strictures directly from ERCP fluoroscopy images[4] and that automated systems trained on multicenter cholangiogram datasets can segment the duct, stones, and the duodenoscope to assess the technical difficulty of stone extraction[5]. Overall, this suggests that fluoroscopic cholangiograms are amenable to computational analysis, as in other domains of endoscopy, although questions remain regarding dataset size, external validation, and clinical readiness[6].

This narrative review focuses on the emerging use of AI in interpreting fluoroscopic cholangiograms during ERCP. We outline the theoretical foundations of the field, briefly summarize current diagnostic, procedural, and predictive applications, and critically evaluate the limitations and validation gaps that impede the translation of promising pilot data into routine clinical practice.

METHODS

This article was planned as a narrative review rather than a systematic review or meta-analysis, given the nature of the available evidence. Studies applying AI to fluoroscopic cholangiography are relatively new and methodologically heterogeneous, spanning diverse clinical settings, neural network architectures, and outcome measures that are not suitable for statistical pooling. To preserve transparency and methodological rigor despite the narrative design, the review was structured according to the Scale for the Assessment of Narrative Review Articles, with particular attention to the rationale for the study question, description of the literature search, and appraisal of the quality of the available evidence[7].

A structured literature search was performed in PubMed/MEDLINE, Scopus, and Web of Science, with support from IEEE Xplore, to capture engineering and computer science articles that may not be fully indexed in biomedical databases. Search strings combined technology terms, including AI, deep learning, machine learning, convolutional neural network (CNN), and radiomics, with clinical context terms, including ERCP, cholangiography, cholangiogram, and fluoroscopy, and relevant target terms, including bile duct, stone, stricture, stent, papilla, cannulation, radiation, ethics, governance, liability, bias, and physician in the loop. The search was limited to full-text English-language publications from January 2018 through July 16, 2026, corresponding to the period during which deep-learning techniques became increasingly used in endoscopic imaging.

Eligible studies applied an AI-based approach to fluoroscopic, cholangiographic, or peri-procedural ERCP imaging for diagnostic, procedural, or predictive purposes. Studies restricted to digital single-operator cholangioscopy images or magnetic resonance cholangiopancreatography (MRCP) were excluded from the primary synthesis and cited only when they provided important comparative or contextual framing, as their source data differ substantially from the projectional fluoroscopic images that are the focus of this review. Reference lists of included studies and relevant review articles were manually searched to identify additional sources.

Instead of pooling results, the evidence was organized qualitatively and by clinical task. For each study, particular attention was given to dataset size, single-versus multicenter provenance, the presence or absence of external validation, real-time feasibility, and overall clinical readiness. This task-focused, appraisal-based approach was selected to communicate not only what AI can accomplish with the cholangiogram but also how close each application is to bedside use.

TECHNICAL FOUNDATIONS: TEACHING MACHINES TO READ FLUOROSCOPY

The fluoroscopic cholangiogram is not a stand-alone component, but one element of an increasingly integrated biliopancreatic imaging workup, in which ERCP and endoscopic ultrasound are deliberately combined to maximize diagnostic and therapeutic yield[8]. Therefore, understanding how AI may interpret the cholangiogram begins with recognizing the computational principles shared across this imaging landscape. Conventional machine learning models rely on features that investigators hand-engineer and select before training, whereas deep learning models ingest images directly and generate their own hierarchical representations, an approach better suited to the variability of medical imaging[6]. CNNs are a widely used deep-learning paradigm in this context; they learn spatial features across successive layers and can be trained to classify an entire image, localize structures within it, or label each pixel. This flexibility makes the projectional cholangiogram, despite its noise and operator-dependent acquisition, a tractable computational target.

Four architectural families are particularly relevant to fluoroscopic interpretation. Image classification networks assign a global label to a cholangiogram, such as malignant vs benign strictures[4]. Semantic segmentation networks, such as U-Net and D-LinkNet, segment structures, including the bile duct, stones, and duodenoscope, pixel-by-pixel and provide the geometric information on which downstream tasks, such as difficulty scoring, depend[5]. Object-detection systems, including those based on transformer-assisted object-detection models, can localize small targets such as the duodenal papilla and cannula in real time during the procedure[9]. Finally, radiomics generates numerous quantitative texture and shape descriptors from a focal region of interest, which can then be fed into predictive models, as applied to the papilla region to predict the risk of complications[10].

Fluoroscopic images have challenges that distinguish them from many other image types used in endoscopic AI, such as white-light endoscopy and cross-sectional imaging. Fluoroscopic images are two-dimensional projections that collapse three-dimensional ductal anatomy, with low intrinsic contrast and frequent superimposition of the spine, endoscope, and dynamically filling contrast medium. Image quality is further influenced by the radiation-dose trade-off of fluoroscopy, patient movement, and substantial variability in acquisition settings across operators and units. These acquisition differences may contribute to transportability gaps. In one biliary-stricture classifier, strong internal discrimination decreased across independent cohorts[4]; this single example should not be generalized to all models or tasks.

Two additional factors determine whether such models can be trusted and reproduced. The first is interpretability. Saliency maps can show which image regions are associated with a prediction, but they do not by themselves establish causal or clinically valid reasoning[4]. The second is data: Deep-learning methods are constrained by the size, quality, and representativeness of their training datasets, while the labor-intensive expert annotation required for fluoroscopic images remains a major bottleneck. Clinical translation also requires representative data, independent external validation, prospective workflow evaluation, and analysis of failure cases[6].

DIAGNOSTIC APPLICATIONS: STONES, STRICTURES, AND DUCTAL ANATOMY

To date, the diagnostic application of cholangiographic AI has primarily focused on characterizing biliary strictures. A deep-learning classifier trained on ERCP fluoroscopy images from three German centers differentiated malignant from benign biliary strictures, achieving a cross-validation area under the receiver operating characteristic curve of 0.89. However, performance decreased to 0.72 and 0.76 in two independent external cohorts[4]. This pattern, strong internal discrimination followed by reduced performance on external validation, reflects the current maturity of the field and frames how the remaining diagnostic evidence should be interpreted.

Comparing these projectional-image results with those obtained through direct visualization is instructive. A meta-analysis of AI applied to digital single-operator cholangioscopy reported a pooled sensitivity of 95% and specificity of 88% for identifying malignancy in indeterminate biliary strictures[11]. This represents strong performance in a different input domain and should not be compared directly with fluoroscopy because patient selection, reference standards, and available imaging information differ.

Stone disease is the second major diagnostic target. The cholangiogram is particularly well suited to automated analysis in this setting because stones, ducts, and instruments are geometrically distinct. A deep-learning system trained on 1954 cholangiograms segmented the common bile duct, stones, and duodenoscope with mean intersection-over-union values of 86.4%, 68.4%, and 95.9%, respectively, providing quantitative measurements of stone size and ductal configuration on which subsequent assessment depends[5]. The lower mean intersection-over-union for stones compared with the duct or scope reflects their small size, variable contrast conspicuity, and frequent partial obscuration by contrast, which is a common challenge in projectional imaging.

Where fluoroscopy-specific evidence is limited, cross-sectional cholangiographic AI provides useful contextual information, although its input data differ substantially. In MRCP, deep-learning models have detected common bile duct stones with accuracy approaching that of radiologists, decreasing from 94% for solitary stones to 70% for multiple stones[12]. Similarly, automated recognition of primary sclerosing cholangitis-compatible ductal change has been performed on three-dimensional MRCP data, with sensitivities and specificities of 95% and 91%, respectively[13]. These studies demonstrate feasibility in different imaging modalities and should not be treated as direct performance benchmarks for fluoroscopic cholangiography.

A unifying theme across these diagnostic applications is that accurate image segmentation provides the foundation for higher-order interpretation. Whether the downstream goal is malignancy prediction, stone quantification, or anatomical mapping, the model must first distinguish the duct, lesion, and instrument from a low-contrast, superimposed background[5]. The technical feasibility of cholangiogram analysis has therefore been demonstrated, but is currently constrained by limited datasets, largely single-task models, and insufficient evidence of robustness for routine clinical application[4,11].

PROCEDURAL AND PREDICTIVE APPLICATIONS: FROM CANNULATION TO COMPLICATION RISK

In addition to diagnosis, the most important clinical utility of cholangiographic AI lies in the procedure itself, beginning with cannulation, a technically demanding step and an important determinant of adverse-event risk. From papilla localization to complication prediction, AI may assist at several procedural touchpoints, as summarised in Figure 1. An early CNN system trained on white-light endoscopic images acquired during ERCP localized the ampulla with a mean intersection-over-union score of 64.1% and classified cannulation as easy or difficult, with recall values of 71.9% and 61.1%, respectively, demonstrating feasibility for image-based localization and difficulty classification[14].

Figure 1 Artificial intelligence-assisted workflow for fluoroscopic cholangiogram interpretation during endoscopic retrograde cholangiopancreatography. Fluoroscopic images undergo quality control, segmentation, and feature extraction before task-specific diagnostic, procedural, or predictive output. A clinical safety gate applies confidence thresholds, out-of-distribution alerts, fail-safe abstention, audit logging, and immediate physician override. The endoscopist retains final interpretive and therapeutic authority and integrates the output with endoscopic, endosonographic, clinical, and procedural information. This original schematic was created specifically for this manuscript and contains no patient material or re-used third-party content. AI: Artificial intelligence; ERCP: Endoscopic retrograde cholangiopancreatography.

More recent transformer-assisted detectors have reported higher object-detection performance on a separate dataset, achieving a mean average precision of 93.2% while also computing the planar distance and direction of movement required for the cannula to access the orifice[9]. The result suggests potential real-time navigation support, but it does not establish clinical efficacy and is not directly comparable with the earlier model because the datasets and performance metrics differed.

After duct access is achieved, AI may guide therapeutic strategy by quantifying the difficulty of stone extraction. The previously described deep-learning difficulty-scoring system translated cholangiographic segmentations into a clinically relevant scale, in which scores of 2 or greater were associated with significantly lower complete-clearance rates (36% vs 86% for lower scores) and a greater need for endoscopic papillary large-balloon dilation[5]. By estimating the likelihood of difficult clearance before instrumentation, such scores may help guide accessory selection, referral decisions, or staged management, and illustrate how a diagnostic segmentation task can directly support procedural planning.

AI-assisted stent-length selection has moved beyond conceptual discussion. In a 2026 model development and validation study, an AI workflow identified 121 of 124 common bile duct strictures and selected a stent length within 1 cm of the reference in 104 of 121 cases (85.95%), with a mean absolute error of 0.81 cm[15]. Radiation exposure was reduced by approximately 202 mGy.cm² per patient, with the greatest improvement among less-experienced endoscopists. These findings support the use of AI as a measurement and selection aid, but not for autonomous device deployment.

Relevant procedural labels also arise from non-artificial-intelligence clinical studies. In 596 patients with common bile duct stones, type III papillary morphology independently increased the odds of difficult cannulation (odds ratio, 2.255)[16]. This association can inform model development and risk stratification, but should not be converted into an automated treatment rule without prospective testing.

A further procedural application is radiation reduction. Evidence remains limited, but the stent-length selection study reported a mean reduction in dose-area product of approximately 202 mGy.cm per patient[15]. This is task-specific evidence from an assistive workflow and should not be interpreted as proof that AI broadly reduces radiation across ERCP.

The predictive dimension is illustrated by post-ERCP pancreatitis, a clinically important adverse event. Researchers developed a radiomics-based predictive model using quantitative features from white-light papillary images. The model reported areas under the curve of 0.825-0.857 across the training, testing, and validation cohorts[10].

García-Marmolejo et al[17] examined the timing of ERCP in acute cholangitis but did not evaluate AI[17]. Its relevance is contextual: An image-only model cannot determine urgency or procedural timing without clinical severity, laboratory, hemodynamic, and resource information. This reinforces the need for multimodal decision support.

Additionally, broader reviews in this area highlight the potential of predictive analytics that combines imaging data with clinical information. Such approaches could help identify patients at higher risk of complications and enable more personalized peri-procedural management. However, these models still require prospective testing and external validation before they can be confidently used in routine clinical practice[18].

The diagnostic, procedural, and predictive tools discussed in this and the previous section share a common developmental pathway. They have evolved from retrospective, single-center proof-of-concept studies toward the real-time, validated integration required for clinical use. Table 1 outlines the clinical tasks, computational methods, best reported performance, validation status, and readiness for implementation of these tools, providing the basis for the critical appraisal that follows.

Table 1 Artificial intelligence applications relevant to fluoroscopic cholangiogram interpretation and peri-procedural endoscopic retrograde cholangiopancreatography imaging.

Clinical task	Imaging input	AI approach	Best reported performance	Dataset and validation	Readiness
Malignant vs benign stricture characterization[4]	Fluoroscopic cholangiogram	Classification CNN + saliency mapping	AUROC 0.89 (internal); 0.72-0.76 (external)	251 patients, 3 centers; internal CV + 2 external cohorts	Experimental-externally validated
CBD stone and duct segmentation[5]	Fluoroscopic cholangiogram	Semantic segmentation (D-LinkNet/U-Net)	mIoU 86.4% (duct), 68.4% (stone)	1,954 cholangiograms, 3 hospitals; internal validation	Experimental-multicenter
Stone-extraction difficulty scoring[5]	Fluoroscopic cholangiogram	Segmentation-derived scoring scale	Score ≥ 2 → 36% vs 86% complete clearance	As above	Experimental-multicenter
Biliary stent-length selection[15]	Fluoroscopic cholangiogram	Segmentation and geometric measurement	104 of 121 selections within 1 cm; mean absolute error 0.81 cm; dose-area product reduced by about 202 mGy.cm2	794 images from 431 patients; model development and validation	Experimental-assistive use only
Ampulla localization and cannulation-difficulty classification[14]	White-light papillary image	Classification/detection CNN	Ampulla mIoU 64.1%; difficulty recall 719% (easy)/61.1% (difficult)	531/451 patients, single center; internal	Experimental-proof-of-concept
Real-time papilla and cannula navigation[9]	White-light papillary image	Swin-transformer detector (4STDH)	mAP 93.2%; outputs cannula distance/direction	1840 images; single dataset + video test	Experimental-proof-of-concept
Post-ERCP pancreatitis prediction[10]	White-light papillary image	Papilla radiomics + machine learning	AUC: 0.825-0.857	2372 patients, 2 centers; multicohort validation	Experimental-multicohort

AI: Artificial intelligence; AUC: Area under the curve; AUROC: Area under the receiver-operating-characteristic curve; CBD: Common bile duct; CNN: Convolutional neural network; CV: Cross-validation; ERCP: Endoscopic retrograde cholangiopancreatography; mAP: Mean average precision; mIoU: Mean intersection-over-union.

DOES AI INCREASE RISK DURING ERCP?

Risk is not determined by procedural complexity alone. It depends on the clinical task, model autonomy, error detectability, reversibility, and downstream consequences. Because ERCP is a high-risk interventional endoscopic procedure in which image interpretation may immediately alter guidewire or cannula positioning, sphincterotomy, dilation, stone extraction, or stent selection, an undetected algorithmic error may have more immediate and less reversible consequences than an error in low-stakes offline image review.

Near-term systems should therefore remain assistive. They should display source images and uncertainty, detect out-of-distribution inputs, abstain when image quality is inadequate, preserve immediate physician override, and record model outputs and final actions. No study identified in this review evaluated autonomous control of biliary radiofrequency ablation or another irreversible therapy.

Evaluation should progress from silent local testing to external validation and prospective human-factors studies, with prespecified stopping rules and outcomes including procedure time, cannulation attempts, radiation dose, technical success, adverse events, override frequency, and near misses.

LIMITATIONS, VALIDATION GAPS, AND FUTURE DIRECTIONS

The most important limitation of cholangiographic AI is the chasm between internal promise and external reality. At present, most studies surveyed are retrospective and single- or few-center in scope, and when independent validation has been attempted, performance has declined. For example, the stricture classifier showed a reduction in discrimination from 0.89 internally to 0.72-0.76 externally[4]. The generalizability of machine-learning models for indeterminate biliary strictures has been specifically questioned, as models calibrated to the imaging features, patient mix, and labeling conventions of selected centers may not translate well to other settings[19]. Until external robustness is demonstrated routinely, internal estimates may overstate performance in new clinical settings.

This is compounded by methodological limitations in studies making clinical claims. The available literature is dominated by diagnostic-accuracy and proof-of-concept studies rather than prospective comparative evaluations. Future clinical trial reports should follow the Consolidated Standards of Reporting Trials-Artificial Intelligence extension, and protocols should follow the Standard Protocol Items: Recommendations for Interventional Trials-Artificial Intelligence extension[20,21]. Early-stage prospective evaluation may follow the Developmental and Exploratory Clinical Investigations of Decision-support systems driven by AI guideline, which emphasizes clinical performance, safety, usability, workflow, and human-artificial-intelligence interaction before large comparative trials[22]. These frameworks improve completeness of reporting and error analysis; they are not themselves risk-of-bias instruments[23].

Several practical challenges remain between current models and bedside use. Fluoroscopic images require manual expert annotation, datasets are heterogeneous across endoscope manufacturers and acquisition environments, and most pipelines have not been designed for real-time intraprocedural use. Large self-supervised foundation models trained on millions of luminal-endoscopy images may improve downstream accuracy and reduce task-specific annotation requirements, but GastroNet-5M is a luminal-endoscopy resource and not a substitute for ERCP fluoroscopy data[24]. Similarly, semi-automated annotation of live colonoscopy video provides only methodological context and requires ERCP-specific, multicenter external validation before translation[25].

Clinical translation will depend on more than model performance alone. The reviewed literature did not identify a regulatory-approved AI system for ERCP as of July 2026, and deployment raises concerns, including false positives, alarm fatigue, automation bias, interpretability, data privacy, and algorithmic bias[6,23]. The most defensible near-term role is assistive, physician-supervised decision support, which still requires prospective safety evaluation. Multimodal integration is a promising future direction: The fluoroscopic cholangiogram may be combined with cholangioscopic, endosonographic, clinical, and laboratory data in an integrated diagnostic and predictive framework that may eventually be supported by foundation-model architectures[23,24]. The central task is therefore validation, standardization, and responsible clinical integration rather than further proof-of-concept demonstrations.

ETHICAL, LEGAL, DATA GOVERNANCE, AND EQUITY CONSIDERATIONS

The World Endoscopy Organization consensus groups the central implementation challenges into data governance, medicolegal implications, and equity and bias[26]. For ERCP, governance should define data stewardship, lawful use, retention, access, de-identification, secondary use, cybersecurity, change control, and incident response. Model, training data, and label provenance should be documented.

The World Medical Association endorses physician oversight of AI in medical care[27]. A licensed physician should review model output and retain final clinical authority, but legal responsibility may also rest with developers, institutions, and other parties according to jurisdiction and control. When AI may materially influence invasive care, patients should receive proportionate information about its role, limitations, data use, and oversight.

Aggregate performance may conceal poorer results in underrepresented groups, altered anatomy, uncommon indications, low-resource settings, or centers using different equipment. Developers should report the dataset composition and the performance of clinically relevant subgroups. Institutions should require local validation, equitable access, and clinician training that preserves the unaided skills required during system failure.

CONCLUSION

AI is an emerging area for interpreting fluoroscopic cholangiograms during ERCP. Current evidence suggests that AI may facilitate differentiation of malignant from benign biliary strictures, segmentation of the bile duct and stones, estimation of stone-extraction difficulty, stent-length selection, papilla and cannula localization, and prediction of post-ERCP pancreatitis. A task-specific assistive workflow has also been reported to reduce radiation exposure. Nevertheless, most studies remain retrospective, small or selected, single- or few-center, and incompletely externally validated; the reviewed literature did not identify a regulatory-approved ERCP system as of July 2026.

A plausible future direction is integration into a multimodal, physician-in-the-loop workflow combining fluoroscopy with cholangioscopy, endoscopic ultrasound, clinical data, and procedural outcomes. Before clinical use, future systems must demonstrate robust external validity, real-time feasibility, explainability, workflow compatibility, and benefits for clinical decision-making and patient-important outcomes. Responsible validation and standardization, rather than proof-of-concept performance alone, should define the next stage of development[23,24]. Current evidence therefore supports transparent, physician-supervised assistance rather than autonomous therapeutic decision-making.