Wireless capsule endoscopes (WCE) are revolutionary devices for noninvasive inspection of gastrointestinal tract diseases. However, it is tedious and error-prone for physicians to inspect the huge number of captured images. Artificial Intelligence supports computer-aided diagnostic tools to tackle this challenge. Unlike previous research focusing on the application of large deep neural network (DNN) models for processing images that have been saved on the computer, we propose a light-weight DNN model that has the potential of running locally in the WCE. Thus, only images with diseases are transmitted, saving energy on data transmission. Several aspects of the design are presented in detail, including the DNN’s architecture, the loss function, the criterion of true positive, and data augmentation. We explore design parameters of the DNN architecture in several experiments. These experiments use a training dataset of 1222 images and a test dataset with 153 images. The results of our study indicate that our designed DNN has an Average Precision of AP25= 91.7 % on our test dataset while the parameter storage size is only 29.1KB, which is small enough to run locally on a WCE. In addition, the real-time performance of the designed DNN model is tested on an FPGA, completing one image classification in less than 6.28ms, which is much less than the 167ms needed to achieve real-time operation on the WCE. We conclude that our DNN model possesses significant advantages over previous models for WCEs, in terms of model size and real-time performance.
Bibliographical notePublisher Copyright:
© 2021, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.
- Deep learning
- Medical image processing
- Polyp detection
- Wireless capsule endoscope (WCE)