Artificial Intelligence model for pancreatic intraductal papillary mucinous neoplasm detection and follow-up on magnetic resonance imaging
2School of Computer Science, The Hebrew University of Jerusalem, Israel
Purpose:
To develop an artificial intelligence (AI) model for automated radiological follow-up of intraductal papillary mucinous neoplasm (IPMN) pancreatic cysts using volumetric assessment on magnetic resonance imaging (MRI).
Methods and Materials:
Seventy-three MRI scans demonstrating IPMN cysts greater than 10mm were selected to serve as the model dataset. The magnetic resonance cholangiopancreatography (MRCP) sequence was used for both manual and automated segmentation of the IPMN cysts. The study was conducted in two phases: in phase one, two expert radiologists determined the IPMN cysts contour manually on the MRCP sequence. In phase two, Neural Network (NN) was trained and optimized for automatic calculation of the the cysts’ volume, based on the automatic segmentation task. For the NN a 3-dimensional (3D) UNet architecture was used that accepts input of 3D patches with data augmentation such as overleap, rotating, and Gaussian noise. The NN outputs 3D patches masks that compose to a 3D mask of the cysts in the scan, using an adaptive thresholding technique. The dice similarity coefficient, a statistical tool to assess similarity between two sets of data, was calculated to compare the network segmentation results versus the radiologists’ manual segmentation.
Results:
The curated data set included MRIs from 67 patients with mean age of 70 years (35-90) with 25M:48F ratio. The data set includes scans with IPMN cysts greater than 10mm that were annotated and verified by 2 experienced abdominal radiologists. The dice similarity coefficient between the neural network result and the expert radiologists’ measurements was 0.79 with a standard deviation of 0.11.
Conclusions:
Artificial intelligence provides a novel method for automatic segmentation and volumetric assessment of IPMN cysts providing a fast, accurate, and consistent cyst evaluation method for disease follow up.