Leveraging artificial intelligence to differentiate benign from malignant biliary strictures: A step toward precision diagnosis

Abstract

Differentiating benign from malignant biliary strictures is a crucial decision-making step in the management of patients presenting with biliary strictures. Despite the availability of various modalities including abdominal imaging and endoscopic interventions, the overlapping features and technical procedural challenges attribute to the difficulty in confirming benign vs malignant strictures. These limitations and the high attributing cost of multimodality diagnostic workups often lead to delays in care delivery and impact prognosis. Artificial intelligence (AI) potentially appears promising in improving diagnostic precision via AI-driven algorithms, including deep learning and radiomics. This narrative review explores the role of AI in the accurate classification of biliary strictures to facilitate early detection of biliary cancer to optimize patient outcomes. Integration of AI into clinical workflows could revolutionize biliary stricture evaluation by offering a non-invasive, consistent, cost-effective and data-driven diagnostic pathway.

Key Words: Artificial intelligence; Biliary strictures; Cholangioscopy; Endoscopic ultrasound; Radiomics; Precision medicine

Core Tip: The current management of biliary stricture is highly dependent on multimodality diagnostic pathways to differentiate the benign or malignant nature of the disease process. Despite the availability of multimodality diagnostic tools, precision in diagnosis remains a high-stake clinical challenge. Various artificial intelligence (AI)-based approaches such as deep learning have emerged as potential tools to enhance diagnostic precision. From lesion characterization to assistance in obtaining cholangioscopy or endoscopic ultrasound-guided targeted biopsies, AI enhanced cross-sectional imaging analysis (radiomics) has a promising role. However, further evidence is required to overcome the limitations including insufficient outcome-driven validation, biases in study designs, and lack of interpretability. Hence, well designed trials, explainable AI systems, integration into existing diagnostic pathways and validation based on outcomes would help to enhance the accuracy and actual utility of AI in precision diagnosis of biliary strictures.

INTRODUCTION

Despite advancements in the diagnostic modalities available to differentiate benign and malignant biliary strictures, a significant proportion of biliary strictures remain undifferentiated; leading to delays in care delivery and impacting prognosis. It is crucial to differentiate malignant biliary stricture due to pancreaticobiliary malignancies such as cholangiocarcinoma, pancreatic ductal adenocarcinoma and periampullary malignancies from benign entities including chronic pancreatitis, iatrogenic injuries, and immunoglobulin G4-related disease for accurate diagnosis and timely evidence-based therapy which can ultimately improve survival[1]. Biliary strictures are classified based on their nature (malignant, benign, and indeterminate) or according to location (distal or hilar) and growth trend (intraductal or extrinsic)[2]. Due to the frequent asymptomatic presentation in these patients, the exact incidence of biliary stricture remains undefined[3]. Generally, an approximate 15%-24% is attributed to benign strictures with iatrogenic biliary injury after cholecystectomy or liver transplant being the most common cause[4]. Up 70% of cases are reported to be malignant biliary strictures with pancreatic adenocarcinomas being the most common underlying cause[1,5,6]. Out of the 57000 annually diagnosed cases of pancreatic cancer, around 60% manifested with symptoms of obstructive jaundice[1]. Cholangiocarcinoma is another common cause of malignant biliary stricture with an incidence of 0.3-6 cases per 100000 population in the West and > 6 cases per 100000 in regions endemic for liver fluke infection. Moreover, patients with primary sclerosing cholangitis (PSC) have a 15% lifetime risk of developing cholangiocarcinoma[7]. Furthermore, around 55% risk malignancy in benign strictures without an apparent mass[1]. The situation becomes even more challenging when evaluating indeterminate biliary strictures, which manifest in up to 20% of cases presenting with narrowing in biliary channels without any sign of malignant transformation on cross-sectional imaging or endoscopy[2]. Although radiologically certain features do help differentiate these lesions such as, benign strictures being mostly small with smooth margins while malignant strictures can be longer with irregular and shouldered margins; however, these features may not be present in all cases[8]. Hence, an immense burden on the healthcare system and self-payers is imposed due to care costs attributed to recurrent hospitalizations and invasive procedures to address the consequences of biliary obstruction[1]. Although missing malignancy results in lost opportunity for timely curative treatment, an incorrect diagnosis of cancer may lead to unnecessary surgery or endoscopic intervention[9]. This phenomenon of the grey zone generates an overwhelming anxiety related to recurrent invasive procedures for both clinicians and patients; emphasizing a critical unmet need in modern biliary medicine. As commonly used modalities such as endoscopic retrograde cholangio-pancreatography (ERCP) with brush cytology and intraductal forceps biopsy have limitations in resolving this diagnostic challenge[10,11], advanced procedures like cholangioscopy and endoscopic ultrasound (EUS) offer excellent visualization; however, their access remains a challenge due to limited availability, operator-dependent interpretation and increased costs[11].

With phenomenal progress in artificial intelligence (AI), especially deep learning and computational capability to analyze histopathological specimens, endoscopic and radiological images, AI has the potential to identify subtle and complex signs of malignancy which are not detectable to the human eye[12]. AI is expected to close this diagnostic gap with data-driven-objective assessment. Hence, in this narrative review we explore the utility of AI in differentiating benign, malignant and indeterminate biliary strictures with the integration of AI functions across different diagnostic modalities, precision medicine frameworks, and highlight challenges for translation into clinical practice.

METHODOLOGY

A comprehensive literature search was conducted via PubMed, Google Scholar, and ResearchGate using search terms such as “artificial intelligence”, “machine learning”, “deep learning”, “radiomics”, “convolutional neural networks”, “computer-aided diagnosis”, “biliary strictures”, “cholangiocarcinoma”, “pancreatic ductal adenocarcinoma”, and “benign OR Malignant OR indeterminate biliary strictures”, “endoscopic ultrasound”, “Endoscopic Retrograde Cholangiopancreatography (ERCP)”, “cholangioscopy”, “cytology”, and “histopathology”.

All related studies, meta-analyses, systematic reviews, basic and translational research illustrating the role of AI in the diagnosis of biliary strictures, published in English and between 1992 and 2026 were reviewed. The reference list of selected articles was also screened to include relevant studies (Figure 1).

Figure 1  Flow chart of the literature search.

Benign or malignant: Why current tools are inadequate?

The current diagnostic algorithm for biliary stricture is inadequate for clinicians. Clinicians essentially need a trustworthy, real-time, and lesion-specific tool which can clearly differentiate benign from malignant disease. ERCP-based sampling such as brush cytology, forceps biopsy and Fluorescence in situ hybridization (FISH) generate incremental but inadequate results as these methods merely examine the superficial epithelium, often failing to detect the submucosal and desmoplastic nature of many cancers[13]. Historically, ERCP was primarily used as a diagnostic procedure. Its function in current practice is largely therapeutic due to the availability of safer diagnostic modalities like magnetic resonance cholangiopancreatography (MRCP) and endoscopic ultrasound (EUS), and the increased risk of complications, notably pancreatitis[10]. Moreover, ERCP can be challenging due to difficult common bile duct cannulation[14]. Cross-sectional imaging plays an important role in the initial evaluation of biliary strictures, particularly for staging and planning a procedural road map. Table 1 summarizes the performance characteristics of key imaging techniques[15-19]. Despite cross-sectional imaging being an essential tool for staging and planning a procedure, it cannot provide histopathologic confirmation of malignancy. In the context of these limitations, a summary of the performance characteristics of commonly used ERCP-based sampling modalities for differentiating between benign and malignant strictures is shown in Table 2[1,11,13,20-26].

Table 1 Cross-sectional imaging for evaluation of biliary strictures.

Imaging	Sensitivity range/specificity range, %	Notes
CT scan[15-17]	Sensitivity 75-80. Specificity 60-80	Study context: Prospective, blinded comparisons in patients with obstructive jaundice. Patient population: Mixed benign & malignant strictures (distal and hilar). Reference standard was histology or > 12 months[16,17]. Primary role: Excellent for staging, vascular assessment, and detecting metastases. Key limitation: Limited specificity for differentiating malignant from benign strictures, particularly in the absence of a discrete mass
MRCP[18]	Sensitivity 83-90. Specificity 94-98	Study context: Prospective observational study comparing MRCP directly to ERCP as a reference standard. Patient population: 60 patients with suspected CBD or pancreatic duct pathologies. Primary role: Non-invasive gold standard for evaluating biliary tree anatomy, stricture morphology, and level of obstruction. High accuracy for choledocholithiasis and ductal dilation. Key limitation: Provides anatomical, not histologic, diagnosis
18F-fluorodeoxyglucose positron emission tomography[19]	Sensitivity 85-92. Specificity 51-90	Study context: Pooled data from a meta-analysis. Patient population: 47 studies (n = 2125 patients). Primary role: Not for primary stricture characterization. High utility for staging (lymph node/distant metastasis). Changes management in approximately 15% of cases, primarily via upstaging. Key limitation: Very low specificity (51%) for diagnosing malignancy at the primary site due to false positives from inflammation (e.g., PSC, cholangitis). Cannot replace histopathological confirmation

CT: Computed tomography; MRCP: Magnetic resonance cholangiopancreatography; ERCP: Endoscopic retrograde cholangiopancreatography; CBD: Common bile duct; PSC: Primary sclerosing cholangitis.

Table 2 Endoscopic retrograde cholangiopancreatography-based sampling modalities.

Baseline modality	Sensitivity and specificity range, %	Combined (with biopsy) sensitivity and specificity range, %	Incremental gain, %	Notes
Brush cytology alone[1,11,13,20-22]	Sensitivity 27-79. Specificity 90-100	Sensitivity 66-83.3. Specificity 95-100	Incremental 15-20 sensitivity over single methods	Data from retrospective cohort (Nanda et al[13] n = 61, CCA diagnosis including hilar, perihilar and distal strictures) and RCT comparing brush types (n = 64, extrahepatic strictures[21]). Data from systematic review and meta-analysis (16 studies); yield improves with multiple passes and bile cytology, RCT shows yield improves with modified biopsy forceps[20,22]
Brush cytology and FISH][1,13,23-26]	Sensitivity 35-84. Specificity 54.1-98	Sensitivity 70-82. Specificity 90-100	20-40 yield gain over cytology alone, highest ERCP yield (triple sampling), real-time risk stratification	Data from retrospective cohort (Nanda et al[13] n = 61, CCA diagnosis including hilar, perihilar, and distal strictures), comparative study by Kipp et al[23]. n = 131, including biliary strictures, data from prospective study, n = 81 with biliary and pancreatic duct strictures[24], data from retrospective study n = 614 patients[25], 10 years retrospective study of 281 patients[26]. FISH detects polysomy (aneuploidy) in 80% malignancies, misses diploid tumors, biopsy has no submucosal access

CCA: Cholangiocarcinoma; ERCP: Endoscopic retrograde cholangiopancreatography; RCT: Randomized control trial; FISH: Fluorescence in situ hybridization.

As described, standard ERCP-based tissue sampling has limited sensitivity for the diagnosis of biliary strictures; hence, adjunctive techniques have been developed to improve diagnostic yield. For example, wire-guided assisted endobiliary forceps (US Endoscopy, Steris, Ohio, United States) have been introduced for controlled insertion of the forceps along a guidewire into the biliary system; refining depth for acquiring tissue specimens[27]. However, prospective comparative studies on peroral cholangioscopy (POCS)-guided biopsy and wire-guided biopsy are lacking, and their role in routine clinical practice is under investigation[28]. One study showed that balloon dilation of biliary strictures can disrupt fibrotic tissue planes, increasing the sensitivity of intraductal biopsy from approximately 41% to 71% without an increased risk in adverse events[29]. Additionally, tube-assisted biopsy techniques in which larger caliber biopsy forceps are advanced in bile ducts have shown improved tissue sampling rates, and are an acceptable alternative when D-SOC is not available or expertise is limited[30]. In addition to these techniques, Singhi et al[31] reported 73% sensitivity and 100% specificity of Targeted Next Generation Sequencing for the diagnosis of malignant biliary strictures when combined with brush cytology. Additionally, intraductal biopsy markedly improved the diagnostic yield of pathological evaluation from 35% to 77% and 52% to 83%, respectively[31]. Moreover, endoscopic scraping devices achieved almost similar tissue adequacy and cancer detection rates to that of POCS-guided biopsy[32].

Intraductal forceps biopsies and brush cytology with advanced techniques are not adequate due to their limited depth of tissue sampling, operator dependency and false negative results. Therefore, advanced technologies such as EUS-guided tissue acquisition, cholangioscopy directed biopsy, intraductal ultrasound, probe-based confocal laser endomicroscopy, and optical coherence tomography allow deeper, precise, and microarchitectural assessment of strictures[1,28].

The next section outlines the performance characteristics of these advanced non-AI modalities that allow more targeted evaluation of indeterminate biliary strictures, as demonstrated in Table 3[28,33-47].

Table 3 Advanced modalities for biliary strictures.

Modality	Sensitivity range, %	Specificity range, %	Advantages/limitations	Notes
Intraductal ultrasound[33-35]	93-97	79-90	Pros: Real-time wall-layer/T-staging, superior for proximal strictures. Cons: Probe insertion issues, limited N-staging, advanced tumors	Large retrospective cohorts (n = 193; n = 379) using histopathology or long-term follow-up as reference standards[34,35]
Cholangioscopy directed biospy[28,36,37]	58-86	90-100	Pros: Direct visualization, targeted sampling in proximal lesions. Cons: Costly equipment, expertise required, distal limitations	Multicenter retrospective SpyGlass DS™ cohort (n = 206)[36]; pooled evidence from systematic review and meta-analysis extracted data from 15 studies (n = 539) by Badshah et al[37]
EUS + FNA[38-40]	76-94	90-97	Pros: Excellent for distal strictures and mass lesions, deeper access. Cons: Perihilar challenges, lower yield without mass. Significant impact on surgical decision-making	Meta-analysis (6 studies, n = 497)[38] and prospective cohorts (n = 50)[39]; n = 44[40] confirm high diagnostic accuracy of EUS-FNA, particularly for extraductal lesions > 1.5 cm. The gold standard was surgery or 6 months follow up
EUS guided FNA + ERCP guided TA[38,39]	86-98	98-100	Pros: Maximizes yield in the same session, reduces false negatives. Cons: Procedural time and risk increase, resource-intensive	Systematic review and meta-analysis of same-session procedures (6 studies, n = 497) demonstrated an accuracy of 96.5%, superior to either modality alone, including hilar, perihilar, and distal strictures[38]. Prospective comparative study (n = 50) showed combined sensitivity 97.9% and accuracy 98%, significantly reducing false negatives[39]. The gold standard was surgery or 6 months follow up
EUS + FNB[41,42]	85-98	90-100	Pros: Better tissue yield than FNA, ideal for distal strictures. Cons: Seeding risk, operator-dependent	Prospective multicenter cohort (n = 465) showed superior tissue core yield and histologic accuracy with 22G FNB (99%) vs FNA (61%), with reduced blood contamination; combined analysis reached 100% diagnostic accuracy[41]
EUS guided FNB and ERCP guided tissue acquisition[43,44,60]	83-100	95-100	Pros: EUS-FNB serves as a first-line or complementary tool for diagnosing indeterminate biliary strictures, offering superior sensitivity (83%-98%) over ERCP sampling, especially for distal/extrahepatic lesions without visible masses. Cons: Increased procedural time/risk, dual expertise needed	Retrospective BS cohort (n = 51) with surgical histology or radiologic/clinical follow-up as gold standard showed higher accuracy with same-session EUS-FNB + ERCP (83.3% sensitivity; 87.5% accuracy; 100% specificity) vs either alone[43]. Sensitivity drops to 56%-75% with stents or hilar location due to access challenges[44]
pCLE[45]	75-80	80-85	Pros: In vivo histology, real-time neoplasia detection, improved with Paris Classification. Cons: Probe fragility, steep learning curve, limited availability, high cost	Validation study of the Paris Classification using 40 pCLE sequences (19 inflammatory, 6 benign, 15 malignant indeterminate biliary strictures) demonstrated improved specificity with maintained overall accuracy (approximately 82%). The gold standard was histopathology or clinical follow-up. Interobserver agreement was fair (κ = 0.37). Lesions involved indeterminate bile duct strictures, primarily the extrahepatic biliary tree during ERCP-based evaluation
OCT/VLE[46,47]	79	69	Pros: Microstructural imaging of desmoplasia excels in tight strictures. Cons: Interpretive variability, limited availability, large RCTs needed to confirm its clinical impact	Prospective ERCP-OCT study (n = 37 biliary strictures; 19 malignant, 16 benign) using histology/EUS-FNA/surgery or ≥ 12-months follow-up as reference standard showed sensitivity 79% and specificity 69% (≥ 1 criterion); specificity 100% when both criteria were required. Combined with brushings, increased sensitivity to 84%[47]

EUS: Endoscopic ultrasound; FNA: Fine needle aspiration; FNB: Fine needle biopsy; ERCP: Endoscopic retrograde cholangiopancreatography; TA: Tissue acquisition; BS: Biliary stricture; pCLE: Probe-based confocal laser endomicroscopy; OCT: Optical coherence tomography; VLE: Volumetric laser endomicroscopy.

Understanding AI

The introduction of AI in medicine refers to a digital system which helps improve efficiency and speed of disease diagnosis and treatment, ultimately improving patients’ outcome[48]. For the diagnosis of biliary strictures, AI offers models particularly machine learning and deep learning, with the latter being an influential technique currently[12]. Machine learning models execute effectively with mixed data sets and have proven fruitful in resource-limited settings. However, they are restricted by interobserver variability in feature selection and potential loss of information from generalization of complex image data. AI models are not manually programmed but trained using clinical data, laboratory results or imaging measurements to automate the decision making process[48]. On the other hand, deep learning employs multiple layers of interconnected neurons to facilitate analysis of complex datasets[48]. In the study by Mascarenhas et al[49], the convolutional neural network (CNN) model illustrated potential for recognizing complex patterns, such as papillary projections, tumor vessels, nodules and masses that could easily be overlooked by the human eye[49]. More recently, vision transformers have emerged as an alternative deep learning model. This employs context aware analysis and self-attention mechanisms between image patches, enabling multimodal data integration with understanding of global context[50]. By integrating this conceptual framework into clinical research, substantial and rapidly growing evidence now demonstrates the use of AI in the diagnosis of indeterminate biliary strictures. Figure 2 clearly summarizes the framework of AI models for the evaluation of indeterminate biliary strictures[51]. AI applications in biliary strictures can be categorized into: (1) Non-invasive, pre-endoscopic models; and (2) Invasive, endoscopy-guided models.

Figure 2 Artificial intelligence frameworks for indeterminate pancreatobiliary stricture diagnosis. CA19-9: Carbohydrate antigen 19-9; CNN: Convolutional neural network; EUS: Endoscopic ultrasound; MRCP: Magnetic resonance cholangiopancreatography.

NON-INVASIVE AI MODELS

AI for non-invasive risk stratification of biliary strictures

AI integrated with radiomics: AI integrated with radiomics is being applied to cross-sectional imaging to address the clinical dilemma of indeterminate biliary strictures, where non-invasive differentiation between malignant and benign etiologies remains challenging. Most contemporary radiomics pipelines rely on supervised machine-learning classifiers or deep learning architectures, particularly CNNs, which automatically extract high-dimensional spatial and textural features from cross-sectional imaging[30]. A study reported a panel of radiomic features for pancreatic ductal adenocarcinomas (PDAC) such as low first order intensity features (mean, median, and 90^th percentile values) and high texture features neighboring a grey-tone difference matrix, busyness, grey-level non-uniformity and dependence non-uniformity. These features capture tumor heterogeneity corresponding to computed tomography (CT) images beyond what is detectable with the human eye with an area under the curve of 0.98 and 0.91; validating that AI captures desmoplastic heterogeneity, showing its importance for indeterminate biliary strictures without clear masses[52]. The process involves manual segmentation and labelling of the CT images to identify region of interest, followed by feature extraction via PyRadiomics version 2.2. Consequently, classification of patches as cancerous or noncancerous using the extreme gradient boosting model[52] is obtained. Model performance is significantly affected by slice thickness, with thin slice venous phase CT imaging demonstrating greater accuracy compared to thick slice arterial-phase imaging[53]. Beyond CT, multidomain features incorporate data from magnetic resonance imaging (MRI) sequences and positron emission tomography (PET)/CT imaging. with imaging phenotypes such as texture features, pixel intensity and shape using fluorodeoxyglucose PET based radiomics correlating with underlying genetic alterations including KRAS and SMAD4, showing that radiomic phenotypes can noninvasively detect tumor genotype and behavior[53-55]. Table 4 provides a summary of radiomics in AI[52-57].

Table 4 Artificial intelligence based radiomics models for non-invasive diagnosis and risk stratification of biliary strictures.

Study year	AI model	Data source	Key performance	Clinical impact
2025, meta nalysis[56]	Various ML, LR, RF, SVM, DL	CT (13 studies), EUS (5 studies), MRI (3 studies), PET-CT (3 studies), 24 case-control studies; 14406 patients (7635 PDAC)	Sensitivity 0.92 (95%CI: 0.91-0.94); Specificity 0.90 (95%CI: 0.85-0.94); AUC 0.94 (95%CI: 0.74-0.99); DOR 110 (95%CI: 62-194)	Non-invasive PDAC screening; reduces EUS-FNA need; CT best for initial triage, limited by moderate radiomics quality score and case-control design
Multi-institutional radiomics study[52]	ML classifiers	CT imaging of PDACs < 2 cm	Sensitivity > 90% for small PDAC detection; cross-population generalizability demonstrated	First to characterize PDAC-specific radiomic signature (decreased intensity, increased NGTDM busyness) linking imaging features to desmoplastic heterogeneity; validated across multiple populations
Differentiation AIP vs PDAC[53].	Radiomics-based ML	Thin-slice venous-phase CT	Accuracy 95.2%; AUC 0.975; sensitivity 89.7%; specificity 100%	Demonstrated superiority of thin-slice venous-phase imaging; proved AI can discriminate malignancy from complex benign inflammatory conditions (previously a major limitation)
PET/CT radiogenomics[54,55]	Radiomic analysis	FDG PET/CT with genomic correlation	Metabolic texture features correlate with KRAS and SMAD4 mutations	First demonstration that radiomic phenotypes reflect underlying tumor genotype; provides non-invasive window into tumor biology beyond structural imaging
2023, narrative review[57]	Radiomics review (various ML)	CT/MRI across HPB studies	PDAC detection/differentiation: AUC 0.71-0.99. Tumor grading prediction: AUC 0.73-0.90. IPMN high-grade dysplasia prediction: AUC up to 0.84	Supports early PDAC screening, cyst risk stratification (e.g., IPMN high-grade dysplasia AUC 0.84), and resectability assessment in pancreatic/HCC/ICC

AIP: Autoimmune pancreatitis; ML: Machine learning; LR: Logistic regression; HPB: Hepatopancreatobiliary; RF: Random forest; SVM: Support vector machine; DL: Deep learning; CT: Computed tomography; EUS: Endoscopic ultrasound; MRI: Magnetic resonance imaging; PET: Positron emission tomography; PDAC: Pancreatic ductal adenocarcinomas; CI: Confidence interval; AUC: Area under the curve; DOR: Diagnostic odds ratio; FNA: Fine needle aspiration; NGTDM: Neighboring grey-tone difference matrix; AI: Artificial intelligence; IPMN: Intraductal papillary mucinous neoplasm; HCC: Hepatocellular carcinoma; ICC: Intrahepatic cholangiocarcinoma; FDG: Fluorodeoxyglucose.

Although cross-sectional imaging-based AI models enable non-invasive risk stratification prior to endoscopy, their true clinical value lies in informing the need for ERCP rather than replacing tissue diagnosis. As illustrated in Figure 3, this has led to the emergence of AI-driven pre-ERCP triage frameworks that integrate radiomic risk scores with clinical parameters to guide decision-making[58].

Figure 3 Artificial intelligence assisted pre-endoscopic retrograde cholangiopancreaticography risk stratification based on discriminative performance. AI: Artificial intelligence; AUROC: Area under the receiver operating characteristic curve; MRCP: Magnetic resonance cholangiopancreatography; EUS: Endoscopic ultrasound; ERCP: Endoscopic retrograde cholangiopancreatography; D-SOC: Digital single-operator cholangioscopy; FNB: Fine needle biopsy.

Pre-ERCP risk stratification is a vital step ensuring administration of ERCP only to patients showing a high likelihood of malignancy; thus curtailing a needless procedural risk and healthcare costs. Current guidelines hold this principle: In cases of PSC, ERCP is not recommended for routine cancer surveillance due to its procedural risks and invasiveness. It should only be performed solely in cases of dominant strictures with high suspicion of malignancy[59]. The European Society of Gastrointestinal Endoscopy guidelines similarly recommend initially performing EUS-guided tissue acquisition (EUS-TA) when drainage is not immediately required, reserving ERCP-TA for cases needing intervention or when EUS-TA yields inconclusive results[60]. However, AI-driven predictive models can effectively identify patients with a high probability of malignancy, minimizing delays in obtaining the optimal treatment strategy.

Why is this not universal? What is the barrier at this specific decision point?

Despite demonstrating promising performance metrics, non-invasive AI models face several stage-specific barriers and challenges that limit universal adoption: At the imaging acquisition stage, considerable variability in protocols is seen across institutions. Models trained in thin-slice venous-phase images can potentially fail when applied to thick-slice or arterial-phase studies; necessitating protocol standardization or robust domain adaptation techniques[53]. During model development, reliance on retrospective single-center datasets introduces selection bias and limits external validity. Most studies employ case-control designs which do not reflect real-world prevalence, potentially overestimating diagnostic performance[56]. Moreover, AI cannot fully replace tissue-based diagnosis, as the overlap between benign-malignant cases persists, emphasizing the need for prospective validation and integration into clinical workflows[61]. This creates a translation barrier for the precise technique to integrate AI predictions into clinical workflows without jeopardizing the gold standard of pathologic confirmation. Real-time AI analysis necessitates seamless Picture Archiving and Communication System integration but does not yet fit naturally into radiologist reporting workflows, thus limiting routine clinical adoptions[61].

INVASIVE AI MODELS

AI integrated with endoscopic evaluation of biliary strictures

AI integrated with ERCP: Emerging AI applications in ERCP are shifting focus toward improving procedural efficiency and safety at the critical point of biliary access. A recent CNN-based system has demonstrated precise real-time detection of the ampulla of Vater and prediction of cannulation difficulty, thereby achieving performance comparable to that of expert endoscopists across a vast array of ampullary morphologies[62]. Such functionality is especially pertinent in cases of biliary strictures, where distorted anatomy, inflammation, or malignant infiltration often render cannulation challenging, and increases the risk of repeated attempts and post-ERCP pancreatitis. By enabling early ampullary localization and anticipating difficulty, AI-based systems may assist endoscopists in implementing appropriate prophylactic measures against post-ERCP pancreatitis, pre-planning procedural strategies, optimizing device selection, and maintaining heightened intra-procedural vigilance ultimately reducing procedural trauma and enhancing the overall safety of therapeutic ERCP. Beyond real-time procedural guidance, artificial neural network (ANN)-based models have demonstrated utility in optimizing ERCP decision-making across both benign and malignant biliary diseases. In patients with suspected choledocholithiasis, Jovanovic et al[63] have illustrated how an ANN integrating laboratory and imaging parameters outperformed multivariate logistic regression [area under the curve (AUC) 0.884 vs 0.787] in selecting patients for therapeutic ERCP, demonstrating that AI can optimize patient selection and reduce unnecessary procedures[63]. More recently, Ma et al[64], applied an ANN model in patients with malignant biliary obstruction undergoing ERCP with stent placement, integrating clinical, laboratory, and tumor-related variables to predict early postoperative mortality. The ANN model demonstrated higher specificity and overall predictive performance compared with logistic regression (AUC 0.813 vs 0.727), indicating that AI can predict procedural risks in malignant strictures to guide patient selection and procedure planning[64]. Recent advances in machine learning have extended beyond procedural planning to the early differentiation of malignant vs benign biliary strictures using ERCP-derived data. Yang et al[65] developed a dual-center, retrospective machine learning model via clinical and laboratory parameters obtained during ERCP. Employing Random Forest classification with explainable AI interpretability, the model identified key predictors including stricture location, stricture length, carbohydrate antigen 19-9 levels, bilirubin profiles, and inflammatory markers The model achieved an area under the receiver operating characteristic curve of 0.988, accurately identifying high-risk patients and guiding clinical decisions[65].

Collectively, the work of Jovanovic et al[63], Ma et al[64] and Yang et al[65] demonstrates that AI in ERCP can serve three distinguishing functions: (1) Refine patient selection for ERCP, ensuring only those who can benefit should undergo the procedure; (2) Anticipate procedural complexity and risk to facilitate pre-procedural optimization and prophylaxis; and (3) Pre-emptively distinguish malignant from benign strictures using routinely available clinical and laboratory data, supporting timely and appropriate intervention.

It is important to recognize that unlike EUS and cholangioscopy, AI in ERCP does not seek to independently diagnose or characterize tissue. Instead, its role is fundamentally supportive by guiding clinical decision-making, procedural planning, safety, and risk mitigation alongside other modalities.

AI in D-SOC

D-SOC allows for direct visualization of the biliary tree but its interpretation remains highly subjective and operator-dependent, compromising diagnostic consistency[66]. AI, particularly CNNs, is being developed to standardize this visual assessment and guide targeted biopsies[49]. A landmark multicenter study recruiting 164 patients developed a CNN to automatically differentiate malignant from benign biliary strictures. Trained on over 96000 D-SOC images, it achieved an accuracy of 92.9%, with a sensitivity of 91.7% and a specificity of 94.4%. Moreover, the same model demonstrated high accuracy in identifying specific visual features of malignancy that are critical for endoscopic diagnosis: Nodules and masses: (93% accuracy), papillary projections: (90.8% accuracy) and tumor vessels: (78.1% accuracy)[49]. Perhaps the most clinically significant innovation is the model’s ability to generate real-time heatmaps. These visual overlays highlight areas with the highest probability of malignancy (e.g., regions with papillary projections or abnormal vessels), directly guiding the endoscopist to the optimal biopsy site, which aims to improve tissue yield and diagnostic accuracy[49]. These findings build on earlier work by Saraiva et al[67] in 2022 in which CNN occlusion heatmaps demonstrated significant spatial concordance with biopsy-positive sites (r = 0.609, P = 0.03), lending mechanistic credibility to AI-driven, real-time biopsy targeting during digital cholangioscopy[67]. In a subsequent expanded study, the same research group utilized 84994 images from 129 D-SOC examinations across two centers in Portugal and Spain to develop a CNN for identifying malignant biliary strictures. This model achieved a diagnostic performance with an AUC of 0.92, enhancing AI’s feasibility in distinguishing benign from malignant findings on cholangioscopy[68]. Contextualizing these performance metrics indicates that the CNN model’s overall accuracy of 92.9% is substantially higher than the Mendoza criteria, which has a reported diagnostic accuracy of 77%, suggesting that AI-driven metrics offer excellent diagnostic performance. The system’s ability to produce these visual explanations serves a dual purpose: It provides actionable procedural guidance while also addressing the ‘black box’ problem inherent to deep learning[12]. Research in other endoscopic domains such as colorectal polyp diagnosis, additionally confirms that higher heatmap coverage of a target lesion correlates strongly with a correct AI prediction, underscoring the value of this transparency for building user confidence and guiding procedural action[69].

Beyond individual procedure optimization, the CNN framework offers strong potential for multimodal integration. This includes combining AI-driven visual analysis with FISH, whole genome sequencing from AI-flagged biopsy sites, and pre-procedural MRCP-based radiomics to generate hybrid nomograms that can improve overall patient outcomes[14]. A particularly valuable contribution of AI assistance is its operator-leveling effect. A multicenter validation study by Robles-Medranda et al[70] demonstrated that the CNN consistently outperformed non-expert endoscopists and even exceeded the performance of one expert, indicating a clear leveling of diagnostic capability. By reducing dependence on subjective visual interpretation and minimizing interobserver variability, AI assistance may shorten the D-SOC learning curve and improve diagnostic consistency, particularly in complex biliary strictures[70]. This effect is especially valuable in challenging proximal or perihilar strictures, where subtle mucosal and vascular patterns are difficult to interpret and traditionally require extensive experience. The technical robustness of these systems is supported by the CNN’s high area under the precision-recall curve of 0.95, confirming resilience to class imbalance, with positive predictive value and negative predictive value both exceeding 90%. Recent advances have further validated this potential. In prospective video analysis, MBSDeiT accurately identified 92.3% of malignant biliary strictures, demonstrating its efficacy in a dynamic, real-time setting with superior performance to that of both expert and novice endoscopists[71]. This performance supports reliable exclusion of malignancy when the model predicts a benign finding. Furthermore, demonstrated inter-center European validation and early United States pilot deployment suggest technical generalizability, positioning such systems favorably for future prospective trials and potential regulatory pathways[49].

AI in EUS

EUS excels in evaluating distal strictures and mass lesions but remains fundamentally limited by high operator dependency, subjective interpretation of ambiguous parenchymal textures, and the inherent risk of sampling error in subtle or infiltrative lesions. CNN based models applied to EUS offer an exemplary shift to overcome human limitations by providing quantitative and objective analysis[14]. The endoscopist can perform fine needle aspiration (FNA) or fine needle biopsy (FNB) in the same setting, and when combined with contrast-enhanced EUS (CH-EUS) and elastography, AI increases sensitivity and specificity for the definitive diagnosis of pancreaticobiliary malignancies[72]. A key advance lies in the AI-assisted CH-EUS MASTER system for real time capture and segmentation of lesions for targeted tissue acquisition. In the prospective study by Tang et al[73] it was reported that AI-guided CH-EUS significantly improved first-pass diagnostic yield compared with conventional EUS-FNA with an accuracy of 0.842 suggesting a promising role of AI in reducing operator dependency and improving sampling efficiency[73]. The integration of AI with contrast enhancement facilitates real-time perfusion mapping and lesion segmentation, effectively transforming biopsy targeting from an art to a science. By objectively identifying the most vascularized or heterogeneous region within a mass for FNA or FNB, AI directly addresses the core issue of sampling error. This results in fewer needle passes, potentially lowering complication rates, with a high likelihood of obtaining a definitive diagnosis on the first procedure, thereby expediting treatment and alleviating patient anxiety. Beyond targeted sampling, AI enhances procedural completeness through real-time anatomical guidance. The EUS-IREAD system is a real-time station recognition with blind spot monitoring which reduces missed stations (4.5% vs 12.3%, P < 0.01); while simultaneously integrating CH-EUS perfusion maps for vascular targeting[74]. This functionality ensures both comprehensive anatomical survey and optimized biopsy site selection within a single integrated platform. AI models also excel at distinguishing between conditions that appear similar on EUS, addressing common diagnostic dilemmas. For instance, CNN-based models analyzing imaging features have demonstrated high accuracy in differentiating autoimmune pancreatitis (AIP) from PDAC, a clinically challenging distinction that traditionally may require steroid trials or repeat procedures[75]. The study by Marya et al[75] created an EUS based CNN-based model with occlusion heatmaps, analyzing still images and video frames in real time, which demonstrated sensitivity and specificity of 90% and 93%, respectively, in differentiating AIP from PDAC, thereby providing timely and appropriate patient care with improved outcomes[75]. Additionally, AI demonstrates a powerful operator-leveling effect in EUS, similar to its impact in digital cholangioscopy. A recent Chinese randomized crossover trial demonstrated that a multimodal AI model incorporating EUS images with clinical data achieved excellent diagnostic performance across both internal and external validation cohorts and improved the diagnostic accuracy of novice endoscopists from 69% to 90% (P < 0.001) when classifying pancreatic[76]. This demonstrates that AI can reduce interobserver variability, shorten learning curves, and expand access to expert-level diagnostic precision not just in D-SOC but across complex pancreatobiliary procedures.

AI in EUS-FNB histopathology

AI completes the precision cascade by analyzing EUS-FNB core biopsies, detecting isolated cancer cells and invasive ductal carcinoma that may be missed by conventional cytology. One model achieved an AUC of 0.984 with 93% sensitivity and 97% specificity, substantially surpassing the performance of rapid on-site evaluation, which typically achieves 81.6% accuracy[77]. Moreover, an EUS-based AI model not only guides targeted biopsy but is increasingly efficient in determining the quality of acquired biopsy specimens, thereby closing a critical quality-control gap in EUS-guided tissue acquisition. This quality-control function extends to the procedure room itself. Two staged CNN-based analysis of EUS-FNB stereomicroscopic images from 96 patients has demonstrated the ability to identify diagnostically relevant core tissue even in blood-contaminated specimens, achieving a sensitivity, specificity and accuracy of 90.34%, 53.5% and 84.39%, as an alternative to expert MOSE for rapid adequacy assessment[78]. Such real-time feedback could enable endoscopists to confirm specimen adequacy at the bedside, reducing the need for repeat procedures and improving resource utilization. When integrated with digital pathology, AI can transform glass slides into digital slides that can be viewed on computer screens by extracting quantitative morphological and spatial features, and provide computer-aided diagnosis. This technology provides pathologists with remote access to pathology slides, faster consultation and improved differentiation of benign vs malignant biliary strictures, ultimately providing diagnostic accuracy from lesion targeting to definitive pathological diagnosis[79]. In a pilot study, a computer-aided detection and diagnosis system was applied to 84 whole-slide images derived from bile duct brushings obtained during ERCP. The AI-assisted workflow maintained diagnostic accuracy while significantly reducing cytologist review time, and AI-alone performance was non-inferior to expert cytopathologists. These findings highlight the potential of AI-enabled digital pathology to standardize cytologic interpretation, reduce interobserver variability, and improve efficiency in ERCP-based tissue diagnosis[80]. Beyond immediate tissue adequacy, machine learning algorithms have demonstrated high reliability in classifying PDAC vs benign tissue, grading tumors, and even predicting molecular subtypes, with F1 scores and AUC values often exceeding 90% across internal and external datasets[81]. This ability to extract prognostic and predictive information from routine histology specimens indicates that AI may eventually enable virtual molecular profiling, guiding personalized treatment decisions without the need for additional specialized testing.

EUS precision cascade

A recent systematic review involving 1465 patients from five studies evaluating AI-assisted endoscopic imaging for malignant biliary strictures and cholangiocarcinoma demonstrated that CNN-based cholangioscopy achieved the highest diagnostic accuracy, while CNN-assisted EUS offered the greatest clinical workflow utility through station recognition and bile duct segmentation, reducing procedure time and providing real-time operator feedback[82]. Collectively, these advanced techniques support a shift from general EUS-AI applications toward specific biliary stricture diagnostic precision cascade, in which AI informs lesion flagging, digital pathology aided targeted biopsy, and next generation sequencing within a combined diagnostic pathway for indeterminate biliary strictures, as demonstrated in Figure 4[74,83].

Figure 4 Endoscopic ultrasound precision cascade. AI: Artificial intelligence; CNN: Convolutional neural network; D-SOC: Digital single operator cholangioscopy; EUS: Endoscopic ultrasound; ERCP: Endoscopic retrograde cholangiopancreatography.

Together, these advanced techniques support a conceptual framework of AI models into non-invasive, radiomics based systems for pre-endoscopic risk stratification and invasiveness, particularly CNN-based endoscopic models, that improved lesion identification with targeted biopsy, ensuring tissue adequacy and procedural safety. This evolving framework supports an AI-driven precision cascade for the evaluation of indeterminate biliary strictures, as demonstrated in Figure 5[49,58,63,64,72,73,78,79].

Figure 5 Artificial intelligence-driven precision diagnostic cascade for biliary strictures. AI: Artificial intelligence; AUROC: Area under the receiver operating characteristic curve; MRCP: Magnetic resonance cholangiopancreatography; ERCP: Endoscopic retrograde cholangiopancreatography; D-SOC: Digital single-operator cholangioscopy; EUS: Endoscopic ultrasound; FNB: Fine needle biopsy.

Limitations and potential mitigation strategies

Despite the favorable diagnostic performance of AI in the characterization of biliary strictures, several key limitations must be addressed before widespread adoption in clinical practice. Currently, the evidence remains mainly single-center and retrospective, introducing substantial risks of selection bias and undoubtedly increasing reported performance metrics relative to what might be achieved in real-world clinical practice[14,71]. Furthermore, the field has focused heavily on technical accuracy metrics: Sensitivity, specificity, area under the curve, while largely neglecting assessment of meaningful downstream clinical outcomes, including impact on procedural decisions, patient morbidity, time to definitive diagnosis, and healthcare utilization[84,85]. Compounding these concerns, static images or pre-recorded videos have been evaluated by AI based models rather than during live procedures. Hence leaving many questions unanswered in real-world settings, thus limiting integration into a conceptual framework, reducing generalizability[86,87]. These limitations are not merely academic; they directly impact the safety, generalizability, and ultimate clinical utility of AI systems for patients with biliary strictures. Table 5 summarizes each major challenge, supporting evidence, and corresponding mitigation strategies[12,14,67,70,71,74,84-89].

Table 5 Limitations of artificial intelligence and potential mitigation strategies in the diagnosis of biliary strictures.

Limitations	Evidence from current literature	Illustration from AI biliary stricture studies	Potential mitigation strategies
Retrospective design and selection bias[14,74]	The majority of published AI studies are retrospective and single-center, limiting generalizability and inflating performance estimates	Zhang et al[71] (MBSDeiT) demonstrated prospective real-time D-SOC prediction on a retrospectively trained model, but did not assess the impact on clinical outcomes or compare performance with biopsy guidance. Saraiva et al[68] similarly acknowledged that despite their large dataset for a proof-of-concept study, clinical validation requires a much larger volume of data	Conduct prospective, multicenter randomized trials and pragmatic validation studies
Black-box decision making[12,68]	Most high-performing CNN and ensemble models lack intrinsic interpretability, reducing clinician trust in AI-guided lesion targeting and biopsy decisions	Saraiva et al[68] demonstrated post-hoc heatmap validation of CNN attention to tumor vessels and papillary projections but lacked real-time interpretability during procedures, similarly acknowledged that despite their large dataset for a proof-of-concept study, clinical validation requires a much larger volume of data	Integrate explainable AI techniques (e.g., Grad-CAM, SHAP) as real-time overlays on D-SOC/EUS. Develop AI systems that highlight and verbalize decision-driving features
Limited dataset size and diversity[70,74,88]	Stricture-specific EUS and cholangioscopy datasets remain small (< 10000 labelled images across studies), with underrepresentation of perihilar and benign inflammatory strictures	Robles-Medranda et al[70] conducted a two-phase multicenter study (n = 164), yet disease spectrum and procedural variability remained limited, with ongoing discrepancy between operators visual impression using current classifications for indeterminate biliary strictures	Create open-access, multi-vendor endoscopic image repositories (e.g., expanding The Cancer Imaging Archive. Employ federated learning to train models across institutions without sharing raw data
Domain shift and device dependency[70,89]	Model performance may degrade when applied across different EUS, D-SOC processors, probes, contrast agents, or imaging protocols	The multicenter validation by Robles-Medranda et al[70] may be affected by variability in D-SOC systems and protocols	Implement cross-platform training, harmonization of acquisition protocols, and external validation across vendors and geographic regions
Lack of clinical workflow integration[86,87]	Many AI models are evaluated offline on static images or pre-recorded videos, without assessment of real-time feasibility, procedural impact, or endoscopist interaction	Only Marya et al[87] implemented and validated real-time D-SOC classification during live procedures	Conduct human-factors studies on endoscopist-AI interaction, and conduct real-time deployment trials measuring procedural metrics
Absence of outcome-driven validation[84,85]	Most studies report diagnostic accuracy but lack data on downstream clinical outcomes (time to diagnosis, avoided ERCP, surgical yield)	A multicenter AI study[85] demonstrated superior diagnostic accuracy but did not report the impact on avoided ERCPs or surgical outcomes	Design endpoint-driven trials linking AI-guided diagnosis to clinical outcomes. Perform formal cost-effectiveness and patient-centered measures

D-SOC: Digital single-operator cholangioscopy; AI: Artificial intelligence; CNN: Convolutional neural network; EUS: Endoscopic ultrasound; ERCP: Endoscopic retrograde cholangiopancreatography; CAM: Grad-CAM: Gradient-weighted class activation mapping; SHAP: SHapley Additive exPlanations.

The limitations mentioned above reveal a persistent gap between technical performance and clinical utility. Statements regarding “superior accuracy” should therefore be interpreted with appropriate caution, as reported metrics have not yet been shown to translate into improved patient outcomes or to generalize across diverse clinical settings. The risk of premature adoption in deploying systems that perform well in research datasets but fail in routine practice must be taken seriously.

CONCLUSION

AI appears promising in classifying benign vs malignant biliary strictures. Its potential role could be aligning AI based diagnostic algorithms with conventional diagnostic to improve both diagnostic precision and procedural efficiency. However, further studies are needed to address the current limitations of AI generated diagnostic algorithms to enhance augmented intelligence, a precision cascade for the generation of precise diagnostic information at the right stage of the disease with accurate diagnostic outcomes.