Towards interactive breast tumor classification using transfer learning

The diagnosis of breast cancer relies on the accurate classification of morphological subtypes in histological sections. Recent advances in image analysis using convolutional neural networks have yielded promising automated methods for this classification task. These networks are usually trained from scratch and depend on hours-long training with thousands of labeled examples to produce good results. Once trained these methods can not easily be adapted in cases of misclassification or to novel tasks. We aim to develop methods that can quickly be adapted in an interactive way. As a first step in this direction we present a classification method that enables fast training with a limited number of samples and achieves state-of-the-art results.6