Endoscopic ultrasound (EUS) with fine-needle biopsy (FNB) remains a cornerstone of morphological verification for subepithelial and pancreatobiliary lesions. However, diagnostic performance varies considerably depending on operator expertise, needle selection, and subjective image interpretation. Across studies published between 2018 and 2024, the sensitivity of EUS-FNB ranges from 68% to 89%, and the proportion of nondiagnostic samples reaches 10–25%, particularly in small, fibrotic, or heterogeneous lesions.
Recent developments in artificial intelligence (AI)- including convolutional neural networks (CNN), radiomics, and automated texture analysis- have demonstrated promising results in improving reproducibility and reducing subjectivity in EUS interpretation. Existing AI-EUS studies show AUROC values between 0.86 and 0.95 for differentiating pancreatic mass types and identifying suspicious echotextural patterns. These advancements suggest that incorporating AI into EUS-guided sampling may enhance the precision and consistency of morphological diagnosis.
Based on the collected evidence, we propose the rationale for developing an AI-supported concept- EUS-Precision- aimed at improving the targeting strategy and morphological yield of EUS-FNB through automated pattern recognition and real-time decision guidance. This is not presented as a ready-to-use system, but as a justified direction derived from existing evidence.
The literature analysis revealed several consistent trends:
• AI-enhanced EUS image interpretation improves interobserver agreement and reduces variability in assessing echogenicity, margins, vascular patterns, and internal heterogeneity. CNN-based models in pancreatic lesion analysis demonstrate classification accuracies of 82-94%, surpassing non-expert interpretation.
• Radiomic feature extraction provides quantitative descriptors (texture, entropy, wavelet-based metrics) associated with malignant transformation in pancreatic solid lesions and gastrointestinal subepithelial tumors. Some studies report radiomic AUROC values of 0.87-0.93 for differentiating GIST from benign mesenchymal lesions.
• Factors associated with nondiagnostic FNB samples include lesion size <15 mm, necrotic/heterogeneous internal architecture, and suboptimal targeting. Up to 20-30% of sampling errors occur due to inaccurate selection of the puncture zone rather than needle type.
• AI-supported guidance systems tested in simulation environments or retrospective datasets improve selection of optimal puncture trajectories, reducing simulated nondiagnostic sampling by 15-22%.
These data collectively justify the feasibility of an AI-supported concept such as EUS-Precision, which would integrate:(1) automated echotexture recognition,(2) radiomic-based risk stratification,(3) real-time highlighting of potentially optimal biopsy targets, and(4) decision-support cues to minimize sampling errors.
AI-enhanced EUS interpretation represents a promising direction for improving the diagnostic performance of EUS-guided tissue acquisition. Based on existing evidence, integrating automated echotexture analysis and radiomic markers may improve reproducibility, reduce nondiagnostic sample rates, and support more accurate morphological verification-especially in small or visually challenging lesions.
The conceptual framework of EUS-Precision aligns with ESGE priorities in digital endoscopy, standardization, and operator-independent diagnostic pathways. In the future, such systems may be incorporated into endoscopic workstations as real-time decision-support tools, enhancing diagnostic precision and offering an educational resource for early-career endoscopists.