Unsupervised representation learning for medical imaging

Abstract

[en] The aim of this project is to achieve good image representations through the use of Deep Learning technologies which are part of a broader family of Artificial Intelligence methods named Machine Learning. These image representations are vectors of representative float numbers called embeddings that, for the project, are focused on the entire digestive aparatus for medical purposes. Triplet loss is used altogether with a state of the art Convolutional Neural Network, ResNet, in order to achieve this goal. A series of tests are done in order to compare different training approaches of the ResNet model, seeking for the best image representations in the domain of the digestive aparatus. It is shown that ImageNet transfer learning underperforms with respect to not applying transfer learning for really specialized domains. To conclude, it is found that unsupervised representation learning through the use of Triplet loss enables transfer learning for specialized image domains such as the digestive aparatus.