When doctors find a lesion on the pancreas, one of the most critical questions is: what kind is it? Pancreatic lesions fall into two main categories — cystic lesions (fluid-filled) and solid lesions — and each category has important subtypes with very different implications. Cystic lesions can be mucinous (with cancer risk) or non-mucinous (largely benign). Solid lesions are most commonly pancreatic ductal adenocarcinoma (P-DAC, very aggressive) or neuroendocrine tumors (P-NET, often more treatable).
Endoscopic ultrasound (EUS) is the preferred imaging method for evaluating pancreatic lesions — it provides high-resolution images from inside the digestive tract and allows tissue biopsy. However, EUS diagnosis depends heavily on operator experience and has poor agreement between physicians. A deep learning tool that could automatically classify lesions from EUS images would reduce diagnostic uncertainty.
The study enrolled 378 patients from four reference centers across two continents: Portugal, Spain, the United States, and Brazil. EUS examinations were recorded as videos, which were then broken down into still images, yielding 126,000 total images for analysis.
Two separate convolutional neural networks (CNNs) were developed. A trinary CNN was trained to differentiate between normal pancreatic tissue, mucinous cystic neoplasms (M-PCN), and non-mucinous cystic neoplasms (NM-PCN). A binary CNN was trained to distinguish pancreatic ductal adenocarcinoma (P-DAC) from pancreatic neuroendocrine tumor (P-NET). Both networks used a ResNet architecture pre-trained on ImageNet, with the final layers replaced and fine-tuned for this task.
The complete dataset was split 90/10 into training and testing sets. All histopathological diagnoses were confirmed by tissue sampling (EUS-FNA/FNB), surgery, or fluid analysis — ensuring the ground truth labels were reliable. The study was approved by ethics committees at all four participating institutions.
The trinary CNN for cystic lesion classification achieved exceptional accuracy: 99.1% for identifying normal pancreatic tissue, 99.0% for mucinous cystic neoplasms, and 99.8% for non-mucinous cystic neoplasms. These results represent a major advance over the roughly 48-94% accuracy range reported for human EUS experts.
The binary CNN for solid lesions distinguished P-DAC from P-NET with 94.0% overall accuracy. Sensitivity was 98.7% for P-DAC and 83.7% for P-NET. The relatively lower sensitivity for P-NET likely reflects the smaller number of P-NET training images available (29,153 vs. 64,286 for P-DAC).
The CNNs used transfer learning — starting from a ResNet model pre-trained on millions of ImageNet photographs and fine-tuning it on EUS images. This approach allows useful image recognition features to be adapted to medical images even with relatively limited medical training data.
The transatlantic, four-center study design was a deliberate choice to minimize demographic bias. Training data from centers in four countries across two continents ensured the model was exposed to patients of diverse ethnicities, different EUS equipment (Olympus GF-UCT180, GF-UC140, and SonoScape EG-UC5T), and varied procedural styles. This makes the results more likely to generalize to new clinical settings.
EUS diagnostic accuracy currently varies widely between expert and non-expert endoscopists. An AI decision-support tool that analyzes EUS images in real time could raise the floor of diagnostic quality — helping less experienced endoscopists perform closer to expert level. This is especially important as EUS becomes more widely available globally.
For cystic lesions, accurate classification is critical because mucinous lesions need careful surveillance and may require surgery, while non-mucinous lesions can usually be observed without intervention. Reducing misclassification could prevent unnecessary surgeries and ensure high-risk patients get timely treatment.
This study developed the first worldwide CNN capable of detecting and differentiating the most common pancreatic cystic and solid lesions in EUS images, using data from four centers across two continents. The high accuracy results — particularly for cystic lesion classification — suggest this tool could have an immediate impact on clinical practice.
The authors acknowledge that larger multicenter validation studies are needed before technology implementation. Future studies should also test performance in real-time EUS procedures (rather than on recorded videos) and explore integration into commercial EUS systems. Extension to rarer lesion types is also a future priority.