Pancreatic cancer is notoriously difficult to detect early because it causes few symptoms until it has grown or spread. When lesions are found incidentally, they may be too small to characterize reliably or ambiguous in nature.
Small solid pancreatic lesions under 2 cm are particularly difficult to characterize. Yet detecting and resecting tumors at this size is associated with dramatically better survival, driving the development of better imaging tools.
Endoscopic ultrasound (EUS) has emerged as the most sensitive tool for detecting and characterizing small pancreatic lesions, providing high-resolution images by placing a probe next to the pancreas via the stomach or duodenum.
Standard EUS imaging has been enhanced by two major adjuncts. Contrast-enhanced harmonic EUS (CH-EUS) involves injecting a contrast agent and using harmonic imaging to detect blood flow patterns within a lesion, as malignant lesions typically have different enhancement patterns from benign ones.
EUS-elastography measures the stiffness of a lesion. Cancer tissue is typically harder than surrounding normal pancreatic tissue or benign cysts, and elastography translates this stiffness into color maps that add functional information beyond the structural image.
Both CH-EUS and elastography have improved the diagnostic accuracy of fine needle aspiration and biopsy, with better characterization before biopsy allowing samples to be taken from the most suspicious regions.
AI — particularly convolutional neural networks and deep learning — is being applied to EUS images to automate the detection and classification of pancreatic lesions, analyzing image texture, echogenicity patterns, and lesion borders in ways beyond subjective human assessment.
Early AI models trained on EUS images have shown promising results for distinguishing pancreatic cancer from chronic pancreatitis — a differentiation that is clinically difficult because both conditions can produce similar-looking masses.
Fractal analysis — a mathematical technique that quantifies the complexity of EUS image patterns — has been explored as an AI-compatible feature extraction method for pancreatic lesions.
Studies reviewed show that AI-assisted EUS analysis achieves diagnostic accuracy for pancreatic lesion characterization that meets or exceeds that of experienced endoscopists, with some studies achieving sensitivities and specificities above 90%.
The combination of standard EUS with AI post-processing has been shown to improve interobserver agreement — different endoscopists reviewing AI-assisted images reach more consistent conclusions than when reviewing standard images alone.
For cystic lesions specifically, AI has shown promise in identifying high-risk features like mural nodules and main duct involvement that predict malignant transformation and trigger surgical referral.
The clinical value of AI-enhanced EUS is clearest for characterizing small solid pancreatic lesions below the detection threshold of CT and MRI, and for risk-stratifying pancreatic cysts to decide which need urgent surgery versus continued surveillance.
For small solid lesions in particular, an AI tool that can reliably distinguish suspicious lesions from benign ones could dramatically change management, currently based on incomplete information.
EUS is an operator-dependent procedure requiring significant training. AI assistance could help less experienced endoscopists achieve performance closer to that of expert centers, extending benefits of high-quality EUS assessment to more patients.
This review concludes that EUS, enhanced by contrast agents and elastography, is the current gold standard for detailed evaluation of pancreatic lesions. The integration of AI represents the next major step — enabling objective, automated characterization that surpasses what any individual endoscopist can achieve.
The field is still maturing: most AI studies use small, single-center datasets and have not been prospectively validated. Building large, multi-institutional EUS image libraries is the next essential step.
As datasets grow and AI models are validated, AI-enhanced EUS could become a routine part of pancreatic cancer screening and diagnosis — helping detect disease at a stage when curative treatment is still possible.