Histological image analysis using deep learning

The classification of high dimensional images is becoming more and more relevant as image processing techniques make ground into fields such as medicine and biology. In particular, these techniques are starting to see an application in the computerized analysis of histological images, to determine, for example, whether a tissue is carcinogenic or not. In this MSc thesis we explore the Multiple Instance Formulation for high resolution image classification and extend it to arbitrary sized images. We also review the state of the art techniques used for sampling patches from high resolution images, and propose a new sampling paradigm which has shown to improve the generalization performance of the trained classifier. We compare the proposed methods on two custom made artificial datasets based on the MNIST dataset, and two histological datasets for breast cancer and sun exposure classification.