Distinguishing pancreatic cancer from non-cancerous masses — particularly a condition called mass-forming chronic pancreatitis — is one of the most difficult challenges in digestive medicine. The two conditions look nearly identical on standard imaging, and misdiagnosis in either direction carries severe consequences: missing a cancer delays life-saving treatment, while incorrectly diagnosing cancer leads to unnecessary, major surgery.
Endoscopic ultrasound (EUS) is currently the most sensitive tool available, but its accuracy depends heavily on the experience of the physician performing it. EUS-guided fine needle aspiration (FNA) — where a needle is passed through the scope to take a tissue sample — is considered the gold standard, but finding the best spot to sample within a complex mass is genuinely difficult, especially without real-time guidance.
The CH-EUS MASTER system was built in two parts. Model 1 uses a deep convolutional neural network (based on a powerful architecture called UNet++) trained on 3,546 expert-annotated images to automatically identify and outline pancreatic masses in real time during the endoscopy procedure. This eliminates the need for the physician to manually locate and track the mass.
Model 2 analyzes the time-intensity curve (TIC) — a graph showing how a contrast agent flows through the mass over time. Because pancreatic cancer has a different blood supply pattern than inflammation, this TIC signature can help identify malignancy. A Random Forest machine learning algorithm was trained on 3,552 TIC recordings from 296 patients to distinguish cancerous from benign masses based solely on these blood flow patterns.
In a prospective clinical trial of 39 patients, CH-EUS MASTER correctly identified 24 of 27 malignant masses and 12 of 12 benign masses. Its accuracy (92.3%), sensitivity (92.3%), specificity (92.3%), and area under the ROC curve (0.923) were all significantly better than the performance of experienced endoscopists, who achieved 87.2% accuracy and 0.865 AUC under the same conditions.
When CH-EUS MASTER was used to guide exactly where to insert the biopsy needle during EUS-FNA, the first-pass diagnostic yield was significantly improved compared to procedures done without AI guidance. The AI-guided group achieved 93.8% diagnostic accuracy on the first pass versus 91.3% in the control group, with no adverse events or complications in either group.
A key innovation of CH-EUS MASTER is that it operates within the existing endoscopy system during the live procedure, rather than requiring separate offline analysis. The AI displays a real-time heat map showing which regions of the mass it classifies as cancerous (red), non-cancerous (blue), or necrotic (gray), giving the endoscopist immediate visual guidance about where to aim the biopsy needle.
This integration addresses one of the core limitations of current practice: endoscopists must mentally estimate the best biopsy site while managing the technical demands of the procedure. By providing real-time, objective guidance, CH-EUS MASTER reduces the skill-dependence of EUS-FNA and could help extend high-quality pancreatic diagnostics to centers without highly specialized endoscopists.
CH-EUS MASTER represents a significant advance in AI-assisted pancreatic diagnosis: the first system to integrate real-time mass segmentation, malignancy classification, and biopsy guidance into a single endoscopy workflow. Its performance exceeded that of experienced endoscopists in all key metrics while remaining safe with zero complications.
The main limitation is the relatively small sample size of 39 patients in the randomized trial portion, which limits the statistical power to confirm the improvement in FNA yield. Larger multicenter trials are needed to confirm these results across different patient populations, equipment brands, and operator experience levels before routine clinical adoption.