Attention Based High Resolution Image Classification

Abstract

Modern digital images, especially in the field of medicine, have extremely high resolutions. Current state-of-the-art image recognition techniques, like Convolutional Neural Networks, cannot handle such high dimensional inputs. In this thesis I compared the standard approach ofclassifying images by downscaling them with an attention-based Multiple Instance Learning approach where the original image is split up into several smaller patches and low dimensional embeddings are calculated for each patch by a Convolutional Neural Network. All low dimensional embeddings are then again processed in a MIL fashion, where attention-pooling is used to determine the importance of each patch. The data set for this thesis consisted of ultra high resolution histological slides of human skin which were classified to contain Basal Cell Carcinoma or not.