International Session（Symposium）10(JGES・JSGE・JSGS・JSGCS)
Sat. November 7th   14:30 - 17:00   Room 11: Portopia Hotel South Wing Topaz
IS-S10-8_E
 Usefulness of Deep Learning Analysis for the Diagnosis of Malignancy in IPMN
Takamichi Kuwahara1, Kazuo Hara1, Yasuhiro Shimizu2
1Department of Gastroenterology, Aichi Cancer Center Hospital, 2Department of Gastroenterological Surgery, Aichi Cancer Center Hospital
Background: It is difficult to diagnose the malignancy of intraductal papillary mucinous neoplasm (IPMN) preoperatively. The aim of this study was to evaluate the diagnostic ability of IPMN malignancy using artificial intelligence (AI).
Methods: Retrospective study was performed on patients who underwent endoscopic ultrasonography (EUS) before pancreatectomy and had pathologically confirmed IPMN. The final diagnosis was defined as benign (low grade dysplasia) and malignant (high grade dysplasia/ invasive carcinoma). As training data, total 3,970 still images from 50 patients (04/2013-12/2017) were used to train the deep learning algorithm and the diagnostic ability was evaluated by using 10-fold cross validation. As test data 30 lesions (01/2018-01/2019) were included and were diagnosed by real-time software using AI. EUS features of IPMN (mural nodules: MN, main pancreatic duct diameter: MPD, and cyst size: CS) were also measured in test data to compare the diagnostic ability of AI.
Results: The final diagnosis (benign/malignant) of train and test data were 27/23 and 14/16. In train data, the diagnostic abilities (sensitivity, specificity, and accuracy) of AI were 95.7/92.6/94.0. In test data, the diagnostic abilities of AI were 100.0/81.3/90.0, and significantly higher than that of all EUS features (MN: 84.6/70.6/76.7, MPD:35.7/81.2/60.0, CS: 57.1/62.5/60.0) (P<0.001).
Conclusions: AI may be more accurate method to diagnose the malignancy of IPMN than the conventional EUS features.
Index Term 1: IPMN
Index Term 2: artificial intelligence
Page Top