DEEP LEARNING CAN ACCURATELY DISTINGUISH BETWEEN LOW GRADE DYSPLASIA AND HIGH GRADE DYSPLASIA/ INVASIVE CARCINOMA IN IPMN BY UTILIZING ENDOSONOGRAPHIC IMAGES

Aims Management of intraductal papillary mucinous neoplasms (IPMNs) is currently consensus based on high-risk stigmata and remains a challenge. While low grade IPMNs could be stable for years favoring a watch and wait concept, high-grade IPMN possess a more timely risk of development into invasive pancreatic cancer. To date however there is no established method to distinguish low grade from high-grade IPMNs/early invasive carcinoma before surgery with an acceptable accuracy. We aimed to develop and validate a deep learning-based computer aided detection system to achieve this distinction in endoscopic ultrasound (EUS) images.

Methods We collected 4084 endoscopic ultrasound images of 55 patients who underwent pancreatectomy in our clinic. All patients had histologically proven IPMN. A convolutional neural network with pretrained weights was fine tuned to classify “low grade IPMN” from “high grade IPMN/invasive carcinoma” in 3355 training images from 44 patients. We evaluated our model on a test set with 729 images from 11 patients with no patient overlap to the training set.

Results The convolutional network classified low grade from high grade/invasive carcinoma in the test set of 729 images with an accuracy of 100%.

Conclusions This pilot study demonstrates that deep learning can predict the grade of dysplasia in IPMN from EUS images prior to surgery with high accuracy.

Publication History

Article published online:
14 April 2022