Out-of-distribution medical image detection in deep learning

Abstract

[en] Over the past few years, deep learning has been the most attractive field of artificial intelligence due to its high performance in a broad spectrum of application areas, such as image classification, speech recognition and natural language processing. However, one of the biggest challenges for neural networks is learning to recognize what they do not know: out-of-distribution inputs. In this thesis, we use a medical image dataset from wireless capsule endoscopy (WCE), a novel endoscopy technique that allows better inner visualization of the gastrointestinal tract of patients. In the medical field, building robust models able to detect out-of-distribution images is even more critical, as these rare images could potentially show severe conditions that should not remain undetected. The inability to detect out-of-distribution inputs is precisely one of the big hurdles that must be overcome before deep learning models can be considered reliable enough to be deployed in the real world. This thesis gives a general overview of neural networks from a mathematical point of view and reviews the state of the art of out-of-distribution image detection. Moreover, we present a novel patch-based approach that allows implementing ODIN on unlabeled data and improves out-of-distribution detection performance with respect to the traditional full-image ODIN.