Background and Aims:
Risk stratification and recommendation to surgery regarding intraductal papillary mucinous neoplasm (IPMN) is currently based on consensus guidelines. Risk stratification from presurgery histology only could potentially be decisive but suffers from the low sensitivity of fine needle aspiration. In this study, we developed and validated a deep learning-based method to distinguish between IPMN with low grade dysplasia and IPMN with high grade dysplasia/invasive carcinoma from endoscopic ultrasound (EUS) images.
Patients and methods:
For training our model, we acquired a total of 3355 EUS images from 43 patients who underwent pancreatectomy from March 2015 to August 2021. All patients had histologically proven IPMN. We used transfer learning to fine tune a convolutional neural network and to classify “low grade IPMN” from “high grade IPMN/invasive carcinoma”. Our testset consisted of 1823 images from 27 patients, recruiting 11 patients retrospectively, 7 patients prospectively and 9 patients externally. We compared our results with the prediction of international consensus guidelines.
Results:
Our approach could classify low grade from high grade/invasive carcinoma in the test set with an accuracy of 99.6% [99.5%,99.9%]. Our deep learning model achieved superior accuracy in prediction of the histologic outcome compared to any individual guidelines, which have accuracies between 51.8% [31.9%,71.3%] and 70.3% [49.8,86.2].
Conclusion:
This pilot study demonstrates that deep learning in IPMN-EUS images can predict the histological outcome with high accuracy.
Keywords: Endoscopic Ultrasound, EUS, IPMN, Deep learning, artificial intelligence