Deep Learning to Classify Intraductal Papillary Mucinous Neoplasms Using Magnetic Resonance Imaging

This study aimed to evaluate a deep learning protocol to identify neoplasia in intraductal papillary mucinous neoplasia (IPMN) in comparison to current radiographic criteria.

A computer-aided framework was designed using convolutional neural networks to classify IPMN. The protocol was applied to magnetic resonance images of the pancreas. Features of IPMN were classified according to American Gastroenterology Association guidelines, Fukuoka guidelines, and the new deep learning protocol. Sensitivity and specificity were calculated using surgically resected cystic lesions or healthy controls.

Of 139 cases, 58 (42%) were male; mean (standard deviation) age was 65.3 (11.9) years. Twenty-two percent had normal pancreas; 34%, low-grade dysplasia; 14%, high-grade dysplasia; and 29%, adenocarcinoma. The deep learning protocol sensitivity and specificity to detect dysplasia were 92% and 52%, respectively. Sensitivity and specificity to identify high-grade dysplasia or cancer were 75% and 78%, respectively. Diagnostic performance was similar to radiologic criteria. Areas under the receiver operating curves (95% confidence interval) were 0.76 (0.70-0.84) for American Gastroenterology Association, 0.77 (0.70-0.85) for Fukuoka, and 0.78 (0.71-0.85) for the deep learning protocol (P = 0.90).

The deep learning protocol showed accuracy comparable to current radiographic criteria. Computer-aided frameworks could be implemented as aids for radiologists to identify high-risk IPMN.