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

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.

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.

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.

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.