Artificial intelligence in endoscopic ultrasound: Clinical translation of a prediction, navigation, and diagnosis framework

Abstract

Endoscopic ultrasound (EUS) remains operator-dependent with notable diagnostic variability. This review synthesizes recent artificial intelligence (AI) advances within an integrated “Prediction-Navigation-Diagnosis” framework to transform EUS practice. Preoperatively, AI aids risk stratification and procedure planning. Intraoperatively, real-time navigation systems reduce anatomical blind-spot miss rates by approximately 10% and guide puncture paths. Postoperatively, AI enhances diagnostic accuracy for various gastrointestinal lesions, and cytology models alleviate reliance on scarce pathological resources. However, clinical adoption faces challenges including data heterogeneity, high costs, ethical ambiguities, and insufficient regulatory frameworks. Future translation depends on standardizing multimodal data, developing accessible algorithms, establishing human-AI collaboration guidelines, and advancing adaptive regulations. Overcoming these barriers may enable AI-enhanced EUS to achieve a more consistent, safe, and accessible intelligent workflow.

Key Words: Endoscopic ultrasound; Artificial intelligence; Computer-aided diagnosis; Real-time navigation; Fine-needle aspiration

Core Tip: This article proposes an artificial intelligence-powered “Prediction-Navigation-Diagnosis” framework for endoscopic ultrasound. It highlights how artificial intelligence improves preoperative risk stratification, reduces intraoperative missed detections (about 10%) via real-time navigation, and enhances diagnostic accuracy. Key challenges for clinical adoption - data heterogeneity, high costs, and unclear regulations - are summarized, with future success hinging on multimodal integration and adaptive policies.

INTRODUCTION

Endoscopic ultrasound (EUS) represents a cornerstone technique for diagnosing and treating digestive diseases, particularly valuable in early pancreatic cancer detection, differentiation of submucosal tumors, and guidance of fine-needle aspiration (FNA)[1]. However, the effectiveness of EUS remains highly operator-dependent, resulting in significant diagnostic variability, subjective image interpretation, and a high missed detection rate in anatomical blind zones. These limitations collectively increase procedural risks and reduce diagnostic consistency. Artificial intelligence (AI) has emerged as a promising solution to these challenges through preoperative planning and risk prediction, real-time navigation for optimized operational workflow, and improved diagnostic accuracy in image analysis. Consequently, AI-enhanced EUS could improve clinical decision-making and reshape current practice. This article proposes an innovative and comprehensive “Prediction-Navigation-Diagnosis” framework, integrating AI into the entire EUS cycle. It systematically reviews the transformative role of AI from preoperative and intraoperative stages to postoperative management (Figure 1). This framework offers a translational roadmap for enhancing the precision and accessibility of EUS-based diagnosis and treatment. Its clinical application pathway can be outlined as follows: (1) AI tools are leveraged to perform preoperative risk assessment and patient stratification, facilitating the development of individualized surgical protocols; (2) During the intraoperative phase, real-time navigation and guidance systems are employed to standardize operational workflows and circumvent high-risk regions; and (3) Multimodal AI analysis is utilized to conduct postoperative diagnosis and complication prediction, thereby establishing a closed-loop management system encompassing decision support, procedural execution, and outcome evaluation. This framework addresses existing gaps in the literature, as previous reviews on EUS have primarily focused on isolated stages[2], and provides a translational roadmap for enhancing precision and accessibility in EUS-based care.

Figure 1 Integrated artificial intelligence framework in the endoscopic ultrasound workflow. This schematic depicts a three-part closed-loop framework, termed “Prediction-Navigation-Diagnosis” (indicated by arrows), through which artificial intelligence (AI) enhances endoscopic ultrasound. Six core application domains (shown in boxes) include AI-driven preoperative prediction (preoperative planning and risk stratification), intraoperative navigation (real-time navigation and puncture path guidance), and postoperative diagnosis (cytological diagnosis and imaging diagnosis). The central pie chart demonstrates that the representative AI models cited in this review predominantly originate from 2024-2025 (over 70%), highlighting the recent and advanced nature of developments in this field.

LITERATURE SEARCH AND SELECTION

Databases

PubMed, EMBASE, IEEE Xplore, and China National Knowledge Infrastructure.

Search strategy

EUS combined with AI/deep learning and Computer-Aided Diagnosis/Real-Time Navigation/FNA.

Inclusion and exclusion criteria

Inclusion criteria: Original studies on EUS-AI models published between 2022 and 2025.

Exclusion criteria: Studies lacking clinical validation or containing duplicate data.

Risk of bias assessment

The risk of bias in included studies was assessed using the Prediction model Risk Of Bias Assessment Tool. Most current evidence originates from single-center, retrospective case series with limited sample sizes and without prospective external validation. Consequently, the overall risk of bias is high, primarily due to non-representative populations, lack of blinding, and absence of validation in independent cohorts. These limitations signify that the reported performance metrics may not generalize to broader clinical settings. It must be acknowledged that the field remains in a relatively early translational stage; high-quality evidence from large-scale, prospective, multicenter trials is still scarce. Therefore, AI-enhanced EUS should currently be regarded as an assistive, rather than a replacement, technology in clinical practice.

CLINICAL CHALLENGES OF CONVENTIONAL EUS

Limitations in diagnostic efficacy

The diagnostic accuracy of EUS depends significantly on the operator, resulting in considerable inter-observer variability in lesion interpretation. For example, the diagnostic accuracy of conventional EUS for gastric stromal tumors ranges from only 45% to 67%[3,4]. Additionally, EUS imaging alone has inherent limitations in differentiating specific lesion types. Specifically, distinguishing benign from malignant solid pancreatic masses based solely on EUS images has limited specificity (50% to 60%), increasing the risk of misdiagnosis[5]. Variations in operator experience also cause procedural blind spots and missed lesions. Critical anatomical regions, such as the pancreatic head and the pancreatobiliary junction, are frequently overlooked, with missed lesion rates ranging from 70% to 93% among operators[6].

Procedural risks

Several procedure-related complications limit the safety of EUS-guided FNA (EUS-FNA). Post-procedural acute pancreatitis occurs in approximately 1.8% of cases, often due to iatrogenic injury of the pancreatic ductal system[7]. Perforation is less common (around 0.4%), influenced by operator inexperience, anatomical variations, and pre-existing conditions such as gastrointestinal strictures or diverticula[8]. Hemorrhage occurs in nearly 2% of procedures, with increased risk when targeting highly vascular organs like the pancreas[8]. Furthermore, inadequate specimen acquisition is a persistent issue, potentially reducing diagnostic yield and necessitating repeat procedures[9].

Imaging limitations

As an EUS-based modality, EUS is susceptible to imaging artifacts, such as reverberation, acoustic shadowing, and speckle noise, which complicate interpretation and require considerable expertise to identify and reduce. Although high-frequency transducers (typically 5-20 MHz) provide high spatial resolution, their penetration depth is limited to a few centimeters. Consequently, deep anatomical structures distal to the probe, such as retroperitoneal portions of large tumors or pancreatic visualization in obese patients, are often inadequately visualized, reducing diagnostic clarity.

In summary, conventional EUS faces significant limitations, including pronounced operator-dependent variability in diagnostic performance, procedure-related complications such as pancreatitis (1.8%), perforation (0.4%), hemorrhage (2%), frequent difficulties in obtaining adequate tissue samples, and fundamental imaging limitations inherent to EUS physics.

APPLICATIONS OF AI IN EUS

Intelligent preoperative planning and risk stratification

EUS can detect biliary sludge and microlithiasis, known risk factors for acute pancreatitis[10]. Given its invasiveness and cost, proper patient selection is essential to avoid both overuse and underdiagnosis. Sirtl et al[11] developed a machine learning model using data from 218 acute pancreatitis patients across three tertiary centers. The model incorporated age, liver enzymes, lipids, and urinalysis to predict the risk of biliary sludge/microlithiasis. It achieved 97.9% sensitivity, capturing nearly all sludge-positive cases, and a negative predictive value of 78.6%, thereby reducing unnecessary EUS in low-risk patients. A web-based tool was also created for EUS triage based on patient metrics[12]. Despite certain limitations, this method balances diagnostic accuracy and efficiency, though future prospective validation is needed.

Endoscopic resection is the first-line minimally invasive treatment for subepithelial lesions (SELs), highlighting the need for precise preoperative risk assessment[13]. Using EUS data from over 1000 patients across eight hospitals in Shandong, Chen et al[14] developed the SEL Endoscopic Resection Predictor. Based on lesion location, size, morphology, and echogenicity, SEL Endoscopic Resection Predictor predicts four clinical outcomes: Ulcer management (standard vs advanced closure), procedure time > 30 minutes, hospital stay ≥ 5 days, and medication costs exceeding hospital averages. Although rare outcomes such as perforation were excluded due to low incidence, larger studies may further expand their predictive capabilities. Such tools support personalized planning, resource allocation, and preoperative counseling.

Real-time intraoperative navigation and puncture path guidance

EUS is crucial for diagnosing digestive tract diseases[15-17]. However, its effectiveness varies widely among operators, with miss rates ranging from 70% to 93% due to anatomical blind spots[6]. Several AI-powered systems have been developed to improve scanning standardization and procedural guidance. Wu et al[18] created EUS-intelligent and real-time endoscopy analytical device, a real-time AI assistant that identifies eight standard scanning positions and 24 anatomical structures, reducing missed detection rates by 9.8% for positions and 11.2% for structures in a 290-patient trial. Independently, Li et al[19] developed EUS-automatic image report system, trained on 230000 images, which improved standard position documentation by 13.3%, achieved a 96% lesion capture rate, and identified puncture needles within one second. Although neither system significantly improved lesion diagnosis, both substantially reduced operator workload and improved examination completeness. Further enhancing navigation, Rizzatti et al[20] created the Endoangel system that combines deep learning with real-time voice and text guidance to identify key areas, automate bile duct measurements, and guide the operator’s actions. It shows potential for standardizing EUS training and shortening learning curves, though larger-scale validation remains necessary.

In summary, these systems exemplify a new generation of AI-augmented EUS platforms: EUS-intelligent and real-time endoscopy analytical device serves as a scanning supervisor, EUS-automatic image report system as an automated recorder, and Endoangel as an anatomical navigator. Future developments may include lesion-specific targeting and puncture path planning to further improve precision and accessibility.

Intelligent diagnosis

Imaging-driven AI diagnostic models: Gastric mesenchymal tumors, such as gastrointestinal stromal tumors (GIST), leiomyomas, schwannomas, and other SELs, are commonly identified during gastroscopy[21,22]. However, their differentiation through EUS is challenging, with conventional diagnostic accuracy ranging from 45% to 67% due to high operator variability[3,4]. To address this issue, Hirai et al[23] developed a deep learning model using EfficientNetV2-L trained on 631 pathologically confirmed SEL cases from 12 institutions. The model achieved an overall accuracy of 86.1% in five-class classification (GIST, leiomyoma, schwannoma, neuroendocrine tumor, and ectopic pancreas) and a sensitivity of 98.8% for GIST detection. However, accurately distinguishing schwannomas remains challenging. A limitation of this model is its current use of static images, without incorporating dynamic EUS features (e.g., elastography, Doppler flow) or clinical metadata. In contrast, Joo et al[24] developed a multimodal fusion model integrating EUS image features with clinical variables (patient age, tumor location) using machine learning (random forest). This model achieved an area under the curve of 0.896 and a sensitivity of 93.8%, significantly improving the identification of GIST cases requiring surgery and providing robust clinical decision support. Together, these developments highlight the potential of unimodal and multimodal AI approaches to enhance diagnostic accuracy and standardize SEL evaluations using EUS.

Subepithelial esophageal lesions, including leiomyomas and GIST, typically follow a benign course, though some carry malignant potential. Considering China’s high esophageal cancer incidence (five-year survival: 15%-25%)[25-27], EUS is crucial for localizing lesions (mucosal, submucosal, muscularis propria)[28-30]. However, EUS diagnostic accuracy depends significantly on the operator[31]. To address this issue, Liu et al[32] developed a dual-module deep learning system (RetinaNet-based locator, VGGNet-based classifier). It achieved 82.49% accuracy, 90.56% specificity, and 99.3% accuracy for determining tumor origin, significantly outperforming conventional methods. Similarly, Zhang et al[33] developed a lightweight “detection-stratification” model (YOLOv8s-seg for lesion contouring, MobileNetv2 for layer identification). This system achieved 92.2% precision in lesion contouring and 76.5% accuracy in distinguishing superficial from deep lesions, comparable to senior endoscopists (74.9%-79.8%) and outperforming junior practitioners (65.6%). Despite promising results, both models require multi-center validation, real-time video analytics, and multimodal data integration to support clinical decision-making effectively.

Solid pancreatic lesions present diagnostic challenges due to ambiguous etiology (benign vs malignant)[34-37]. Although EUS is standard for morphological evaluation, conventional interpretation based solely on EUS imaging features has limited specificity (50%-60%)[38]. Cui et al[39] developed a multimodal AI system (Joint-AI) combining deep learning-based EUS image analysis with 36 clinical parameters (age, symptoms, lab markers). In a cross-validation trial involving 12 endoscopists, this model improved junior endoscopists’ diagnostic accuracy by 21% (69%-90%) and sensitivity by 30% (61%-91%). Senior endoscopists initially skeptical of AI assistance showed increased adoption when given interpretable decision support (visual highlights, explanatory annotations). Future implementation requires addressing clinical integration challenges and establishing clear accountability for AI-assisted diagnostics.

Distal cholangiocarcinoma is a rare but highly aggressive malignancy with nonspecific early symptoms and poor five-year survival (10%-40%)[40,41]. Orzan et al[42] developed the first EUS-based AI model specifically for biliary tract evaluation (non-pancreatic), using data from 112 distal cholangiocarcinoma patients and 1248 EUS images. The dual-module system included a customized convolutional neural network for classification (97.82% accuracy, 100% specificity) and a segmentation module (DeepLabv3+, ResNet50 backbones), achieving 90% accuracy in bile duct segmentation and improved tumor margin identification. Although significant, further large-scale, multi-center validation and real-time video integration are needed for clinical adoption.

Colorectal cancer is the third most diagnosed malignancy worldwide[43]. Unlike colonoscopy, which is limited to superficial evaluation, EUS allows detailed visualization of deeper tumor structures (adenomas vs carcinomas)[44,45]. In a multicenter study involving 554 patients and 8738 EUS images, Men et al[46] created an AI model based on the ResNet50 architecture for triple-classification (carcinoma, adenoma, benign tumors). The model achieved 80.93% accuracy, 72.88% sensitivity, and 84.44% specificity for malignant tumors, outperforming senior endoscopists (κ = 0.674 vs κ = 0.557) and substantially surpassing junior endoscopists (κ = 0.284). This demonstrates AI’s potential to standardize colorectal tumor classification and reduce experience-based variability.

In summary, AI-augmented EUS significantly improves diagnostic accuracy across various gastrointestinal conditions (gastric tumors, esophageal lesions, pancreaticobiliary diseases, colorectal neoplasms; Figure 2). Currently, most AI models face the following common challenges in real-world clinical application: Model performance is highly dependent on the quality of training data and annotation consistency, and diagnostic accuracy may significantly decrease in cases of poor image quality, atypical lesion presentation, or rare diseases. Particularly noteworthy is that static image-based analysis methods fail to fully integrate the incremental information provided by dynamic EUS imaging modalities, such as elastography and doppler flow imaging. Furthermore, the generalizability of models still requires systematic validation in real-world clinical environments involving multicenter, prospective designs, diverse acquisition devices, and operators. Future research should focus on developing AI systems capable of integrating dynamic imaging sequences, adapting to various hardware platforms, and maintaining robust performance in complex clinical scenarios, such as coexisting inflammation and tumors, and assessment of post-treatment changes.

Figure 2 Distribution of artificial intelligence diagnostic models across the digestive system. This figure summarizes representative artificial intelligence diagnostic models for endoscopic ultrasound, organized by anatomical location within the digestive tract. For each model, four key elements are presented: Model (grey background) specifies the algorithm architecture; data modality (purple) indicates the type of input data used; key metrics (blue) report the primary performance indicators; and application (orange) describes the diagnostic task. The models demonstrate the application of Artificial Intelligence across diverse gastrointestinal pathologies, including subepithelial lesions, tumors, and specific malignancies. AI: Artificial intelligence; AUC: Area under the curve; CNN: Convolutional neural network; CT: Computed tomography; dCCA: Distal cholangiocarcinoma; EUS: Endoscopic ultrasound; GIST: Gastrointestinal stromal tumor; MRI: Magnetic resonance imaging; MRP: Magnetic resonance pancreatography; SEL: Subepithelial lesions.

Cytology-driven AI diagnostic models: Although imaging-based AI has significantly improved diagnostic precision, histopathological assessment via EUS-FNA remains the gold standard[47-49]. Its utility, however, is limited by a scarcity of pathologists, as nearly 80% of institutions cannot perform rapid on-site evaluation. This shortage results in diagnostic inconsistencies, especially in equivocal cases such as distinguishing inflammatory changes from early malignancies[50,51]. To address these challenges, Fujii et al[52] developed a transformer-based rapid on-site evaluation-AI system trained on 4059 cytological images, achieving an accuracy of 88.2%, a sensitivity of 91.4%, and an area under the curve of 0.954 through geometric transformation-based augmentation. Fang et al[53] proposed a complementary semi-supervised convolutional neural network, reducing manual annotation by 80% while attaining an internal accuracy of 95.1% and a Kappa of 0.853 for diagnosing challenging cases. Despite these advances, both systems require further validation through multicenter studies and improved clinical integration. Future efforts should aim to integrate cytology-based AI models with EUS imaging to develop unified systems that support puncture planning and real-time diagnostic feedback, thereby decreasing reliance on specialized pathology resources.

AI performance in complex clinical scenarios: While existing AI models excel in typical cases, their generalizability in complex scenarios - such as co-existing lesions, rare tumors, and discrimination between inflammation and malignancy - remains to be validated. For instance, identifying early pancreatic cancer against a background of chronic pancreatitis[39] often leads to a significant drop in AI sensitivity. Furthermore, current models struggle to distinguish between histologic subtypes with similar EUS features, such as neuroendocrine tumors G1 vs G2[23]. Future efforts should focus on constructing multicenter databases of complex cases and developing few-shot learning and uncertainty quantification methods to enhance AI robustness in edge-case scenarios.

Prediction of postoperative complications

Post-EUS-FNA pancreatitis occurs in approximately 1.8% of procedures, with a similar incidence (2%) reported for hemorrhage[7,8]. Current risk prediction primarily relies on clinical scoring systems, such as those provided by the American Society for Gastrointestinal Endoscopy and European Society of Gastrointestinal Endoscopy guidelines, which offer limited granularity for personalized assessments[54,55]. Although no dedicated AI prediction model specifically targets EUS-FNA-related complications, Brenner et al[56] developed a machine learning tool for predicting post-endoscopic retrograde cholangiopancreatography pancreatitis. This model includes 20 variables, such as patient age and procedural complexity[56], and provides real-time risk stratification while simulating the effects of preventive interventions (non-steroidal anti-inflammatory drugs, hydration, stenting) on individual risk reduction. Future research could integrate procedural metrics (number of needle passes, needle type), patient biomarkers (baseline amylase levels), and AI-quantified imaging characteristics (vascularity patterns) to develop accurate prediction models for complications like pancreatitis and bleeding after EUS-FNA.

In summary, AI technologies now support the entire EUS workflow within a multidimensional framework. To facilitate comparative evaluation, Table 1 systematically summarizes representative AI models across four key domains: Preoperative planning, intraoperative navigation, imaging-based diagnosis, and cytology-based diagnosis, including details on algorithm architecture, dataset size, performance metrics, feature types, and validation methodologies.

Table 1 Representative artificial intelligence models for the entire endoscopic ultrasound workflow.

Application field	Ref.	Algorithm type	Sample size	Validation method	Application scenarios	Accuracy (%)	Sensitivity/specificity (%)	PPV (%)	NPV (%)	F1 score	AUC (95%CI)	Public availability/ reproducibility
Pre-procedural	Sirtl et al[11], 2023	Machine Learning	218 patients	Multicenter retrospective	Biliary sludge risk screening	84	97.9/92	89.5	98.2	0.93	0.96 (0.92-0.98)	Code: Not available; data: Not shared
	Chen et al[14], 2024	Predictive model	> 1000 patients	Multicenter prospective	Endoscopic resection outcome prediction	NA	NA	NA	NA	NA	NA	Model: Not released; data: Institutional
Intra-procedural	Wu et al[18], 2023	5 DL models ensemble	290 patients	RCT	Anatomical landmark tracking	NA	NA	NA	NA	NA	NA	Code: Not available; system: Proprietary (EUS-IREAD)
	Li et al[19], 2025	Automated reporting system	114 patients	Prospective trial	Standard site documentation	90.3	NA	NA	NA	NA	NA	Code: Closed source; data: Not public
	Rizzatti et al[20], 2025	DCNN	550 cases	Technical validation	Real-time anatomical navigation	NA	NA	NA	NA	NA	NA	Not publicly available
Image diagnosis	Hirai et al[23], 2022	EfficientNetV2-L	631 cases	Multicenter retrospective	5-class SEL differentiation	89.3	98.8/67.6	85.4	94.2	0.91	0.94 (0.90-0.97)	Code: Not shared; data: Multi-institutional (restricted)
	Joo et al[24], 2024	Random forest	NA	Single-center	GIST intervention prediction	89.6	93.8/81.8	88.9	89.5	0.91	0.896 (0.86-0.93)	Code: Not available; data: Not disclosed
	Liu et al[32], 2022	RetinaNet + VGGNet	NA	Single-center	Depth/origin classification	82.5	80.2/90.6	87.1	85	0.83	0.88 (0.84-0.92)	Not publicly released
	Zhang et al[33], 2025	YOLOv8s-seg + MobileNetv2	NA	Single-center	Layer identification	76.50	NA	NA	NA	NA	NA	Code: Not available; model: Proprietary
	Cui et al[39], 2024	Joint-AI	12 physicians	RCT	Solid lesion diagnosis	90.0	91.0/88.0	89.5	89.8	0.9	0.93 (0.89-0.96)	Code: Not shared; data: Anonymized, available on request
	Orzan et al[42], 2024	CNN + DeepLabv3+	112 cases (1248 images)	Single-center	dCCA detection and segmentation	97.8	100.0/94.4	96.8	100	0.98	0.99 (0.97-1.00)	Not publicly available
	Men et al[46], 2025	ResNet50	554 cases (8738 images)	Dual-center	Identification of 3 colon tumors	80.9	72.9/84.4	78.5	79.8	0.75	0.85 (0.81-0.89)	Code: Not released; data: Multi-center (restricted)
Cytology diagnosis	Fujii et al[52], 2024	Transformer	45 cases (4059 images)	Augmentation test	Malignant cell detection	88.2	91.4/85.0	86.7	90.1	0.89	0.954 (0.93-0.98)	Code: GitHub; data: Partially open
	Fang et al[53], 2025	SSCNN	NA	Semi-supervised framework	Cytology analysis with limited labels	95.1	93.8/97.3	96.5	94.8	0.95	0.97 (0.94-0.99)	Code: Not available; model: Internal use only

This table summarizes key artificial intelligence applications within a “Prediction-Navigation-Diagnosis” framework for endoscopic ultrasound, covering preoperative planning, intraoperative navigation, and postoperative image/cytology diagnosis. Performance metrics (e.g., positive predictive value, negative predictive value, F1 score, area under the curve) are included where reported to facilitate comparative evaluation of clinical validity. The “public availability/reproducibility” column indicates access to code, data, or models, highlighting current transparency challenges. Most cited studies are from 2022-2025, reflecting the rapidly evolving nature of this field. DL: Deep learning; RCT: Randomized controlled trial; DCNN: Deep convolutional neural networks; SEL: Subepithelial lesions; GIST: Gastrointestinal stromal tumors; CNN: Convolutional neural networks; dCCA: Distal cholangiocarcinoma; SSCNN: Semi-supervised convolutional neural network; PPV: Positive predictive value; NPV: Negative predictive value; F1: F1 score (harmonic mean of precision and recall); AUC: Area under the curve; CI: Confidence interval; EUS-IREAD: Endoscopic ultrasound-intelligent and real-time endoscopy analytical device; NA: Not available.

CONSIDERATIONS FOR TECHNICAL IMPLEMENTATION AND CLINICAL INTEGRATION

Reproducibility and rigorous validation

The clinical translation of EUS-AI crucially hinges on reproducibility and rigorous validation. Widespread adoption requires open-source initiatives sharing code, datasets, and model weights, supported by platforms like the Medical Imaging and Data Resource Center. A major translational bottleneck is the frequent lack of independent external validation for published models. Most studies are single-center and retrospective, limiting generalizability across diverse populations and equipment. This validation gap directly undermines real-world reliability and delays clinical integration, as performance often drops outside controlled development settings. Therefore, prioritizing transparent, multi-institutional validation frameworks is essential for building trust and facilitating deployment.

Hardware and software integration

Most advanced AI systems (e.g., Endoangel) rely on specific endoscopic platforms (such as Olympus), which limits their adoption in healthcare institutions with heterogeneous equipment. To achieve practical vendor-agnostic integration, several technical pathways can be pursued: (1) Developing digital imaging and communications in medicine structured report standards-compliant modules for standardized EUS data export; (2) Implementing middleware to convert proprietary data streams into a unified format; and (3) Adopting privacy-preserving distributed learning techniques such as federated learning, enabling multi-center model training without sharing raw data. These concrete steps enhance interoperability and reduce dependency on single-vendor ecosystems.

Cost-benefit analysis

The high initial investment in AI systems (hardware, software licensing, maintenance) and potential operational complexity are major barriers to adoption in resource-limited settings. To strengthen the argument for hospital administrators, a structured health economic framework should be outlined (Figure 3). This framework should encompass not only upfront expenditures but also long-term savings from increased diagnostic yield, enhanced workflow efficiency due to shortened learning curves, and potential reductions in medicolegal risks. Widespread implementation should therefore be supported by clear cost-benefit studies demonstrating the comprehensive healthcare economic value of AI.

Figure 3 Health economic framework for artificial intelligence-enhanced endoscopic ultrasound workflow. This flowchart outlines the procedural integration of artificial intelligence in endoscopic ultrasound, accompanied by a structured health economic framework. it details direct costs (e.g., maintenance, cloud services) and indirect costs (e.g., staff training), while highlighting potential savings from reduced repeat procedures, improved diagnostic yield, shortened learning curves, and lower medicolegal risks. This framework supports the cost-benefit analysis essential for the adoption of artificial intelligence in clinical settings. AI: Artificial intelligence.

Workflow redesign

Embedding AI tools into existing EUS workflows may alter operational habits and increase initial procedure time. Successful integration requires optimizing human-machine interfaces to ensure that AI-assisted prompts are timely, intuitive, and do not interfere with the primary operational field of view, as well as providing training to reduce resistance to use.

Learning curve and training

Traditional EUS training involves a long learning period. AI navigation and diagnostic support systems have the potential to serve as standardized training tools, shortening the time required for novices to reach competency (e.g., reducing from over 50 procedures to fewer). However, this must be accompanied by the design and validation of effective AI-assisted training curricula and evaluation systems.

CHALLENGES IN CLINICAL TRANSLATION AND FUTURE DIRECTIONS

Addressing data bottlenecks: Standardization and multimodal integration

EUS data are characterized by significant heterogeneity arising from variations in probe frequency, inconsistent imaging annotation protocols, and operator-specific techniques. These discrepancies lead to data imbalance, limit model generalizability, and potentially introduce regional biases; for example, models trained primarily on European data may exhibit reduced sensitivity when applied to Asian populations[57]. To address these issues, data standardization efforts should adopt established frameworks like the European Health Data Space. This initiative promotes unified EUS data quality standards, including anonymized image-sharing protocols, to enhance model robustness and minimize bias.

A critical yet often overlooked dimension of EUS is its ability to capture dynamic physiological and mechanical properties - such as tissue stiffness, vascular flow, and contrast kinetics - which are essential for differential diagnosis. Current AI models predominantly rely on static images, thereby losing valuable functional information embedded in sequences like elastography (strain ratio), contrast-enhanced time-intensity curves, and doppler flow signatures[19]. Future standardization protocols must emphasize the inclusion and annotation of dynamic features, along with developing methods to effectively encode temporal and functional changes into AI architectures. This shift from static to dynamic modeling is key to unlocking EUS’s full diagnostic potential.

Beyond standardization, multimodal integration combining EUS with computed tomography, magnetic resonance imaging and clinical data represents a promising strategy to improve diagnostic accuracy[58-60]. For instance, Seza et al[61] developed a multimodal AI model incorporating EUS, contrast-enhanced computed tomography, T2-weighted magnetic resonance imaging, and magnetic resonance pancreatography (magnetic resonance pancreatography). This model increased accuracy in classifying pancreatic cystic neoplasms from 91% to 99%, enabling near-perfect discrimination between serous and mucinous tumors. Future research should prioritize optimizing multimodal fusion methods, uncovering biologically relevant correlations, and extending this diagnostic paradigm to other gastrointestinal malignancies.

Enhancing accessibility and cost-effectiveness: Lightweight models and human-AI synergy

The adoption of advanced AI systems in EUS, such as Endoangel, remains limited by dependency on specific hardware platforms (e.g., Olympus endoscopes). High implementation costs further restrict accessibility, especially in resource-limited settings[62]. Another challenge is the prolonged learning curve for conventional EUS training, typically requiring over 50 supervised procedures. To improve affordability and scalability, future development should emphasize lightweight, vendor-agnostic AI architectures with reduced hardware requirements and operational expenses. Therefore, to enhance affordability and scalability, future development must prioritize the creation of lightweight, vendor-agnostic AI architectures that systematically reduce hardware reliance and operational expenditures.

In this context, hardware compatibility and integration complexity emerge as critical barriers to widespread implementation. Currently, most AI-based navigation systems (e.g., Endoangel) remain tied to proprietary endoscopic platforms (such as Olympus) and depend on high-performance workstations or cloud services, which further escalates deployment and maintenance costs. To overcome this bottleneck, the field should promote open architecture designs and cross-platform adaptation interfaces, encouraging device manufacturers to adopt standardized dataoutput protocols compliant with medical information standards such as digital imaging and communications in medicine, thereby lowering integration barriers. Additionally, a “cloudedge collaborative” computing framework could be explored, where lightweight models are deployed on local devices for real-time navigation and preliminary analysis, while complex computations, model optimization, and largescale training are handled in the cloud. This approach would maintain analytical performance while significantly optimizing the overall cost structure.

In addition to cost-efficiency, AI’s fundamental promise lies in its ability to narrow experience-based performance gaps, effectively “elevating non-expert practitioners to specialist-level proficiency” while optimizing resource use through the avoidance of low-yield procedures[62]. Human-AI collaboration has already shown substantial potential. The Joint-AI model introduced by Cui et al[39] significantly improved junior endoscopists’ diagnostic accuracy (69%-90%) and sensitivity (61%-91%) for solid pancreatic lesions. Furthermore, interpretable decision support increased AI acceptance among specialists[39]. Fang et al’s semi-supervised[53] convolutional neural network demonstrated similar potential by reducing annotation dependence by 80% (using only 20% labeled data) while maintaining accuracy above 93.9% in diagnosing pancreatic ductal adenocarcinoma. However, critical challenges persist regarding conflict resolution when AI and clinician judgments diverge. Multidisciplinary consensus frameworks may provide structured mechanisms to reconcile such discrepancies. Further efficiency could be achieved through active learning strategies, where AI proactively identifies diagnostically challenging or atypical cases for prioritized expert review, creating a continuously improving collaborative workflow.

Ethical and governance framework: Accountability, transparency, and informed consent in AI

Clarifying accountability frameworks: Integrating AI into EUS poses complex ethical challenges, particularly regarding accountability for diagnostic or navigational errors, such as missing critical lesions, where responsibility among clinicians, hospitals, and developers remains unclear. Although systems like EUS-AIRS enhance traceability by documenting procedural details (e.g., spatial relationships between needles and lesions)[19], a systematic framework for accountability has yet to be established.

Transitioning from problem to principle, the establishment of such a framework must be grounded in a clear hierarchy of duties. Clinicians, as the ultimate arbiters of clinical judgment, bear the primary responsibility for decisions made with AI assistance. This necessitates not only operational oversight but also the preserved professional autonomy to critically evaluate and, when necessary, override algorithmic recommendations. Concurrently, healthcare institutions carry an institutional duty to rigorously validate AI tools and ensure their adherence to established clinical standards. Developers, for their part, are obligated to ensure the safety, efficacy, and fundamental transparency of their algorithms, thereby furnishing the trustworthy substrate upon which clinical decisions are made. Ultimately, the crystallization of these distinct yet interdependent obligations requires robust regulatory articulation, as envisaged in instruments like the EU AI Act, which must precisely delineate the scope of liability for each stakeholder[63,64].

Enhancing algorithmic transparency and interpretability: Demystifying the “black-box” nature of AI is essential for building trust. Explainable AI techniques should be promoted to provide visual justification for AI diagnoses (e.g., heatmaps highlighting suspicious areas). Concurrently, adopting “Model Cards” -standardized documentation for each clinical AI model - can disclose details such as training data composition, performance metrics, known limitations, and target populations, analogous to drug package inserts[65,66].

Updating informed consent processes: Traditional consent forms often fail to address the specifics and risks of AI-assisted decision-making. The informed consent process should be revised to explicitly inform patients of AI’s role in their care, explaining its function (assistance rather than replacement), potential benefits (e.g., improved consistency), and uncertainties (e.g., limited judgment in rare cases). In the future, generative AI could assist in creating personalized consent documents and patient education materials, thereby improving communication efficiency and transparency[67-69].

Establishing a robust and adaptive regulatory framework

China currently lacks a comprehensive regulatory framework for clinical AI applications. To address this, a multi-tiered oversight system should be established by drawing upon and synthesizing key strengths of international models (Table 2). This system should include: (1) Adoption of a “Predetermined Change Control” pathway, similar to the Food and Drug Administration model, allowing predefined updates (e.g., adjustments in diagnostic thresholds) without full reapproval, analogous to iterative software upgrades[70]; (2) Implementation of risk-stratified oversight aligned with the European Union’s Medical Device Regulation, requiring rigorous clinical validation for high-risk applications (e.g., AI-assisted early cancer screening) while streamlining evaluation for lower-risk tools (e.g., automated reporting systems)[71]; and (3) Continuous performance monitoring mechanisms, following frameworks developed by organizations like the Japan Gastroenterological Endoscopy Society. These mechanisms would mandate systematic documentation of discrepancies between AI recommendations and clinician decisions, alongside periodic reporting of model performance to maintain accountability throughout the AI lifecycle[72].

Table 2 Comparison of regulatory approaches to artificial intelligence-assisted endoscopic ultrasound in foreign countries.

Feature	United States (FDA)	European Union (MDR/IVDR + AI Act)
Regulatory core	Function-based (what the software does) tiered regulation	Device risk + AI system risk dual matrix regulation.
Classification logic	CADt → CADe → CADx → CADa, with increasing risk	Class I, IIa, IIb, III (device risk), plus the AI Act’s “high-risk” category for most medical AI
Primary pathways	510(k) (substantial equivalence), De Novo (novel low-moderate risk), PMA (high risk)	Self-certification (class I low risk), notified body conformity assessment (class IIa and above)
Adaptability	Predetermined change control plans allow for iterative updates to cleared algorithms within defined bounds without new submission	Update processes under regulations are stricter; significant software changes typically require re-assessment/notification to the notified body
Clinical evidence	Emphasizes prospective, multi-center clinical trials to demonstrate safety and effectiveness	Emphasizes clinical evaluation with comprehensive technical documentation and performance validation, aligned with GDPR
Transparency	Requires disclosure of algorithm performance	The AI act emphasizes transparency, requiring high-risk AI systems to provide clear usage information and ensure outputs are interpretable and overseen by humans
Process flowcharts	AI medical device concept → determine intended use and function: CADt, CADe, CADx, or CADa → assign FDA risk class → generate clinical and technical evidence → submit to FDA for review → FDA approval clearance → post-market surveillance	Medical AI product → determine device risk class per MDR/IVDR → determine AI system risk level per AI act (typically “high-risk”) → comply with both sets of requirements → undergo conformity Assessment by a notified body (for class IIa and above) → obtain CE marking

This table summarizes the fundamental differences between the United States Food and Drug Administration and European Union [Medical Device Regulation/In Vitro Diagnostic Regulation + artificial intelligence (AI) act] regulatory approaches for AI in endoscopic ultrasound. Understanding these distinctions in regulatory philosophy, classification logic, and approval pathways is crucial for the global clinical translation and integration of AI-enhanced endoscopic ultrasound devices. FDA: Food and Drug Administration; MDR: Medical Device Regulation; IVDR: In Vitro Diagnostic Regulation; AI: Artificial intelligence; 510(k): Premarket notification; PMA: Premarket approval; De Novo: Pathway for novel, low-to-moderate risk devices; GDPR: General data protection regulation; Notified Body: Independent organization designated by an European Union country to assess conformity of medical devices; CADt: Computer-aided detection for triage; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; CADa: CompUTER-aIDED aUTONOmous diagnosis; CE: European conformity.

Clinical integration pathway: From framework to workflow

Although the “Prediction-Navigation-Diagnosis” framework provides a clear conceptual structure for AI-enhanced EUS, its clinical translation still faces integration barriers. To facilitate a safe and structured transition from “framework” to “workflow”, we propose a phased, stepwise integration roadmap (Table 3). This roadmap is designed to progressively validate and implement AI tools according to their risk level, ultimately enabling seamless and effective AI integration across the entire EUS workflow.

Table 3 Phased clinical integration roadmap.

Phase	Primary goals	Key evidence required	Regulatory milestones
Phase 1 (1-2 years)	Low-risk auxiliary tasks (e.g., automated measurement, structured reporting)	Feasibility studies, human-machine interaction safety data	Registration or simplified clearance (e.g., class II medical device)
Phase 2 (3-5 years)	Moderate-risk diagnostic support (e.g., lesion detection and classification)	Multicenter diagnostic accuracy studies, clinical utility assessment	Moderate-risk device approval (requires performance monitoring plan)
Phase 3 (> 5 years)	High-risk autonomous functions (e.g., puncture path planning, real-time complication alert)	Large-scale RCTs: Real-world effectiveness data	High-risk device approval (requires a full post-market surveillance system)

This table outlines a risk-escalating, phased pathway for integrating artificial intelligence into endoscopic ultrasound practice. Each phase defines primary objectives, key study designs for validation, and corresponding regulatory milestones, supporting a systematic and evidence-based approach to clinical implementation. RCT: Randomized controlled trial.

This integrated workflow is operationally outlined in Figure 4. It depicts the closed-loop, AI-augmented EUS procedure, beginning with real-time AI navigation and image acquisition, followed by joint AI-physician analysis and decision-making, and culminating in structured reporting. This visual roadmap concretizes the stepwise human-AI collaboration essential for translating the conceptual framework into daily practice.

Figure 4 Closed-loop workflow of artificial intelligence-augmented endoscopic ultrasound procedure. This flowchart delineates the operational integration of artificial intelligence (AI) into the endoscopic ultrasound procedure. Beginning with AI-assisted navigation and image acquisition, it proceeds through sequential steps of AI analysis, physician review, consensus decision-making (with multidisciplinary consultation if needed), and final structured reporting. The diagram visually encapsulates the synergistic human-AI collaboration required to implement the “Prediction-Navigation-Diagnosis” framework in clinical practice. AI: Artificial intelligence; EUS: Endoscopic ultrasound.

Emerging technologies: Multimodal large language models and augmented reality

Beyond the phased integration pathway, the future intelligence of EUS will also benefit from the convergence of emerging technologies such as multimodal large language models and augmented reality. Multimodal large language models can integrate EUS images, clinical notes, pathology reports, and even genomic data to enable automated, structured reporting and provide differential diagnostic support for complex cases, thereby reducing documentation burden and improving decision consistency. Meanwhile, combining real-time AI detection with augmented reality navigation could overlay key anatomical structures, lesion contours, and planned puncture paths directly into the endoscopist’s visual field, enabling “what-you-see-is-what-you-treat” precision and further shortening the learning curve while enhancing procedural safety. These technological advances are expected to drive EUS toward a fully integrated, highly autonomous intelligent surgical platform in the long term (beyond phase 3).

CONCLUSION

In conclusion, AI demonstrates significant potential to enhance EUS through a comprehensive “Prediction-Navigation-Diagnosis” framework, addressing key limitations in operator dependency and diagnostic variability. Current applications in preoperative risk stratification, real-time guidance, and diagnostic augmentation show promising improvements in procedural consistency and accuracy. However, widespread clinical translation faces substantial challenges, including data heterogeneity, implementation costs, and unresolved ethical and regulatory considerations. Future success will depend on standardizing multimodal data integration, developing accessible and robust algorithms, establishing clear governance for human-AI collaboration, and fostering adaptive regulatory policies. By systematically addressing these barriers, AI-enhanced EUS can evolve towards a more intelligent, standardized, and accessible paradigm in precision endoscopy.

ACKNOWLEDGEMENTS

The authors would like to express their sincere gratitude to the colleagues who provided valuable insights during the conceptual development of this review. We also thank the research assistants for their support in the preliminary literature search and data organization.