Out of distribution image detection

Abstract

[en] In recent years, with the evolution of technology and artificial intelligence, the field of medicine has undergone a paradigm shift, and other sciences have been introduced directly into medicine. In the case of this project, we work with a set of images, obtained by a wireless capsule endoscopy, which the patient swallows, and transmits images of the intestines to an external device. This set of images has already been treated, but an important element in disease prevention is the existence of bleeding, blood or other elements in the intestines. The aim of this project is to automate the search for strange elements in the images taken by these wireless cameras, in order to save work for doctors. To do this, an out-of-distribution image detection algorithm is applied. In other words, by training a neural network, it is determined whether an image belongs to the same type of images with which the network has been trained, or not. This is done by a modification of the original classification function of the network, applying a threshold, which determines whether the image belongs to the training distribution or not, in addition to a pre-processing of the images. In this report, a description of the functioning and structure of the neural networks is made, making a section on the convolutional neural networks, which are the most used for image treatment, and therefore are the ones used in this project. The classification method implemented is described, as well as its configuration and the results obtained with the given implementation.