REAL-TIME COMPUTER AIDED DETECTION OF SOLID FOCAL PANCREATIC MASSES IN ENDOSCOPIC ULTRASOUND IMAGING BASED ON CONVOLUTIONAL NEURAL NETWORKS

Aims Endoscopic ultrasound (EUS) imaging has a high accuracy for detection of solid focal pancreatic masses. However, the learning curve to master EUS is prolonged. The aim of our pilot project was to develop a real-time deep learning system used to detect and differentiate solid focal pancreatic masses as compared to normal pancreas.

Methods In this pilot study, deep learning algorithms for localization and segmentation were trained and optimized taking into consideration the trade off between performance and speed, to: 1) find pancreas/tumor in frames; 2) label them; 3) compute their bounding box with the corresponding coordinates and 4) segment them, by producing a mask, which gives pixel-wise segmentation of the pancreas/tumor.

Results 50 patients with normal pancreas or solid focal pancreatic masses were included in the study, with 15 images selected for each patient from the movies stored on the embedded hard disk drive of the ultrasound system. A total of 750 images and their ground-truths were used for training and testing of deep learning segmentation models presented in this study, reaching an average precision of 91%.

Conclusions Our model showed a robust classification of normal pancreas versus solid focal pancreatic masses. The mode has potential to be transferred to real-time EUS imaging. Preliminary evidence suggests that these observations have the potential to improve operating characteristics of EUS by enabling targeting biopsies of focal pancreatic masses and to shorten the learning curve of trainee endosonographers.

Publication History

Article published online:
14 April 2022