Evaluation of AI for the diagnosis of pancreatic cancer in linear EUS : Preliminary results of a SFED multicentre study

Le résumé fourni par la source

Aims EUS is the gold standard to diagnose pancreatic cancer. However, the sensitivity and specificity of EUS is not optimal due to variable echogenicity from one patient to another. On the other hand, some infiltrative forms without hypoechoic mass are difficult to visualize. What's more, the operator can be taken for a ride due to a lack of experience. Finally, fatigue can sometimes lead to poor diagnosis of pancreatic tumors. The aim of our study was to evaluate the performance of AI for the diagnosis of pancreatic cancer in linear EUS. Methods This is a collaborative project with the mathematics laboratory specializing in AI. A pilot phase, carried out with a database of around a hundred linear EUS images of pancreatic pathology, enabled us to assess the value of a Transfer Learning algorithm for diagnosing pancreatic cancer with a diagnostic performance of close to 75%. In order to optimize the AI, a large database of linear EUS images was generated by 9 endoscopy centers in France. Patients were divided into 2 groups: group C=cancer (adenocarcinoma, endocrine tumor, metastasis) and group NC=non-cancer (normal pancreas and non-cancer pancreatic pathologies). All images were anonymized, checked and annotated. The primary endpoint was the diagnosis of pancreatic cancer confirmed by pathological findings and patient follow-up. Results 609 patients were included, 209 in the tumor group (C) and 400 in the non-tumor group (NC). In the C group, 179 had adenocarcinoma, 26 had endocrine tumor and 4 had metastatic renal cancer. In the NC group 156 had a cystic pancreatic lesion (IPMN, mucinous or serous cystadenoma), 70 had acute biliary pancreatitis, 55 had chronic pancreatitis, 11 had autoimmune pancreatitis, 78 had normal pancreas and 30 had other pancreatic pathology. A total of 8,545 images were collected, including 2,975 in the C group and 5,520 in the NC group. The sensitivity, specificity and accuracy of AI for the diagnosis of pancreatic cancer in linear EUS were 80%, 90% and 86% respectively. Conclusions AI appears to be effective to diagnose pancreatic tumors in linear EUS Publication History Article published online: 15 April 2024 © 2024. European Society of Gastrointestinal Endoscopy. All rights reserved. Georg Thieme Verlag KG Rüdigerstraße 14, 70469 Stuttgart, Germany

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.