Out-of-distribution detection in dermatology

Deep neural networks are able to diagnose skin diseases from images accurately, but they still have a long way to go to be trusted in healthcare. In this thesis, we tackle two well-known problems in dermatology using out-of-distribution detection with deep neural networks. First, we show that by making small changes to the neural network model, we can make it capable of doing multi-class classification and novel-class detection at the same time. Then, we show that by combining lesion detection, lesion segmentation, and outlier detection modules, it is possible to build a pipeline capable of identifying suspicious ugly duckling signs from total body photography images.