Analyse des images de capsule endoscopique.

Abstract

Wireless capsule endoscopic (WCE) is a non-invasive device, introduced in 2000, that is used by physicians in the diagnosis of diseases of the gastrointestinal (GI) tract. Despite these limited resources, one capsule can operate for several hours and transmit tens of thousands of images. However, these images are of low quality and frequency. In addition, this tool will not be able to visualize the entire GI tract, as the battery life is limited. The main goal of the present work is to develop algorithms for the automatic processing of these images, to help in the diagnosis and in the reduction of the energy consumption. As a first contribution, we propose an automatic classification of lesions and digestive organs in WCE images. This classification is achieved by using two learning techniques such as learning from scratch of a proposed CNN and a transfer learning of pre-trained CNNs. In a second step, we present a new classification method to automatically detect different diseases of the (GI) tract. It is a deep learning algorithm based on features concatenation of two pre-trained convolutional neural networks. In a second contribution, we propose an intelligent compression scheme, which addresses the energy limitation issues of WCE. The principle is to include a classification feedback loop, based on deep learning, to determine the importance of transmitted images. This classification is used in conjunction with a predictive compression algorithm to intelligently manage the limited energy of the capsule. The goal of such a system is to increase the battery life or to obtain high quality images in specific areas. Based on the results obtained, we conclude that our system is efficient and provides good energy optimization of the capsule.