Neural network analysis of EUS images to differentiate between pancreatic malignancy and pancreatitis

Abstract

Background: The differentiation of focal pancreatitis and pancreatic adenocarcinoma is problematic and often resolved only by pancreaticoduodenectomy. EUS is the most sensitive imaging modality for both conditions, yet ultrasonic criteria for distinguishing the two have not been described and differentiation remains difficult. The aims of this study were to develop a self-learning computer program that can analyze EUS images and differentiate malignancy from pancreatitis, and to compare results obtained with this system with EUS interpretation by experienced endosonographers.

Methods: Twenty-one patients with pancreatic cancer and 14 with focal pancreatitis were included. The diagnosis was confirmed histologically in all cases and each patient had undergone EUS. A single EUS image from each procedure was used for computer analysis. The results were compared with the EUS diagnosis reported at the actual procedure as well that of an endosonographer who reviewed videotapes of the procedures.

Results: The software program differentiated focal pancreatitis from malignancy with a maximal 89% accuracy. With sensitivity set at 100% for malignancy, the program was 50% specific and accuracy was 80%. Sensitivity and accuracy of the endosonographer's impression at the time of EUS were, respectively, 89% and 85%. A sensitivity of 73% and accuracy of 83% were achieved with blinded interpretation of EUS videotapes.

Conclusions: Analysis of EUS images with computer software programs is feasible and compares favorably with human interpretation. The application of this technology to EUS and other imaging scenarios could be a useful adjunct to diagnostic endoscopy and warrants further investigation.