|
Pancreatic Cancer (PC) is one of the most aggressive cancers, with a mortality rate of 98%. Although the diagnosis of PC is difficult in early stages, several imaging techniques support the screening process, i.e, ultra- sonography (US), computed tomography (CT), and endoscopic ultrasound (EUS). EUS procedure reports the highest sensitivity (up to 87%) and histological samples may be acquired during the same procedure. However, EUS sensitivity depends on the gastroenterologist's experience. The presented method performs an automatic frame-by-frame detection of PC in complete EUS videos. First, the images are preprocessed to rearrange the radial image intensities, filter out the Speckle Noise, and perform a contrast enhancement to highlight relevant echo patterns. Then, a pre-trained Convolutional Neural Network (CNN) is adapted to the ultrasound domain by a transfer learning strategy to characterize and classify EUS images between PC and non-PC classes. Finally, mislabeled images are corrected by a temporal analysis. The methodology is evaluated using a data set of 66,249 frames from 55 EUS cases. 18 patients are from PC class and 37 for non-P class. A cross-validation scheme is applied seven times to evaluate the performance of three convolutional neural networks: GoogleNet, ResNet18, and ResNet50 architectures. Best results were 93:2 ± 4:0, 87:7 ± 5:4, 95:0 ± 5:6, and 87:0 ± 6:7 in accuracy, sensitivity, specificity, and F-score, achieved with the ResNet50 architecture.
|
