S239 Revolutionizing Early Pancreatic Cancer Detection: EfficientNetB0 Embedded in a Universally Accessible, Multiplatform Diagnostic Tool for Mobile Cancer Detection and Equitable Cancer Care
Suchith Boodegere Suresh, Ramya Elangovan, Jansi Rani Sethuraj, Kavin Elangovan, Khutaija Noor, Elangovan Krishnan · The American Journal of Gastroenterology · 2025
Introduction: Endoscopic ultrasound (EUS) is the gold standard for evaluation and lesion sampling of pancreatic lesions. In this context, there is a need to differentiate each type of solid and cystic lesions. Additionally, in patients with a diagnosis of intraductal papillary mucinous neoplasms (IPMNs), the determination of grade of dysplasia is crucial for guiding clinical management. However, EUS diagnostic accuracy is still suboptimal. This multicenter study aimed to develop and validate convolutional neural networks (CNNs) to characterize solid and cystic pancreatic lesions and stratify IPMNs into high-grade dysplasia/carcinoma (HGD/C) and low-grade dysplasia (LGD). Methods: This multicenter study included EUS images from 14 centers in Portugal, Spain, the United Kingdom, Brazil, Argentina and the United States of America. Solid lesion differentiation (pancreatic ductal adenocarcinoma vs pancreatic neuroendocrine tumor) was achieved using 63,795 images from 161 patients. A second convolutional neural networks for distinction between mucinous and non-mucinous cystic lesions was trained on 202,002 images from 110 patients. Finally, a third model used the subset of IPMNs (n = 51,046 frames, 30 patients) to distinguish between HGD/C and LGD. Histological confirmation was used as the gold standard. The datasets were split into training (70%), validation (20%), and testing (10%) subsets. Performance metrics included sensitivity, specificity, accuracy, and area under the precision-recall curve. Results: For solid lesions, pancreatic ductal adenocarcinoma was identified with 98.9% sensitivity, 90.0% specificity and 95.8% accuracy, while pancreatic neuroendocrine tumor was detected with 97.8% sensitivity, 97.1% specificity and 97.4% accuracy. Mucinous cystic lesions were identified with 98.5% sensitivity, 88.6% specificity and 95.7% accuracy, while non-mucinous cystic lesions were detected with 98.4% sensitivity, 98.7% specificity and 98.6% accuracy. The model distinguished IPMNs with HGD/C from those with LGD with 97.5% sensitivity, 94.3% specificity and 89.8% accuracy. Conclusion: This is the first multicenter and interoperable study showcasing a holistic approach to focal pancreatic lesions, identifying those with higher malignant potential. To our knowledge this is the first model capable of stratifying IPMNs into LGD and HGD/C, potentially reducing unnecessary surgeries and ensuring timely intervention for high-risk patients. Real-time implementation of these solutions could significantly improve decision-making in EUS and optimize patient outcomes.