Artificial intelligence for pancreatic cyst dysplasia grading: a multicenter endoscopic ultrasound study.
Authors
Affiliations (13)
Affiliations (13)
- Department of Gastroenterology, Centro Hospitalar Universitário de São João, Porto, Portugal.
- WGO Gastroenterology and Hepatology Training Center, Porto, Portugal.
- Faculty of Medicine, University of Porto, Porto, Portugal.
- Department of Gastroenterology, Hospital Universitario Puerta de Hierro Majadahonda, Madrid, Spain.
- Division of Gastroenterology, New York University Langone Health, Long Island, New York, USA.
- Department of Gastroenterology and Hepatology, Hospital Universitario Ramón y Cajal, Madrid, Spain.
- Universidad de Alcalá, Instituto Ramón y Cajal de Investigación Sanitaria, Madrid, Spain.
- Centro de Investigación Biomédica en Red de Enfermedades Hepáticas y Digestivas, Madrid, Spain.
- Department of Gastroenterology and Hepatology, Hospital Universitario Marqués de Valdecilla, Cantabria, Spain.
- Department of Gastroenterology, Gastrointestinal Endoscopy Service, Hospital das Clínicas da Faculdade de Medicina da Universidade de São Paulo, São Paulo, Brazil.
- Hospital Vila Nova Star, São Paulo, Brazil.
- Department of Mechanical Engineering, Faculty of Engineering, University of Porto, Porto, Portugal.
- MD Anderson Cancer Center Madrid, Madrid, Spain.
Abstract
Pancreatic cystic lesions (PCLs) are increasingly detected because of the widespread use of imaging techniques. Among them, mucinous PCLs carry a higher malignancy risk, with intraductal papillary mucinous neoplasms (IPMNs) being the most frequent subtype. Accurate stratification based on the degree of dysplasia-low-grade dysplasia (LGD) versus high-grade dysplasia or carcinoma (HGD/C)-is essential to guide clinical management and avoid unnecessary surgical interventions. This study aimed to develop and evaluate a deep learning model for stratifying IPMNs into HGD/C and LGD using endoscopic ultrasound (EUS) images. This multicenter study included EUS images collected from 5 centers across Spain, Brazil, and the United States. Ground truth classification of IPMNs was established through cytologic and biochemical analysis of cyst fluid, EUS-guided through-the-needle biopsy, or surgical specimens. A deep learning model was trained to distinguish LGD from HGD/C. Model performance was assessed on the basis of sensitivity, specificity, accuracy, and area under the precision-recall curve. A total of 51,046 EUS images were extracted from 30 examinations performed at 5 centers in Portugal, Spain, Brazil, and the United States. The model distinguished IPMNs with HGD/C from those with LGD with a sensitivity of 95.7%, a specificity of 88.7%, and an overall accuracy of 87.2%. The area under the receiver operating characteristic curve was 0.951. To our knowledge, this is one of the first studies to evaluate the potential of an artificial intelligence model for dysplasia grading of IPMNs. Prospective validation of our model is necessary to ensure clinical benefit.