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

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.