A few shot learning approach to classify breast histology images

Abstract

Breast cancer screening starts with clinical evaluation and if an abnormality is detected the patient might undergo a tissue biopsy. Then, a pathologist take stained tissue samples from suspected breast areas. The tissue gets analysed under a microscope and a diagnosis gets established. This study introduces a few shot learning approach to classify breast histology images. We use a pre-trained convolutional neural network model to extract features from histology images and a prototypical network to classify the query embeddings into eight classes that represent the exact type of abnormality. To train the network, we have chosen to use the episodic training method. We experiment by changing the number of ways and the number of shot and we conclude that the model that uses 2-shots 2-ways yields the best classification accuracy, i.e. 92%, on a subset of the Break His dataset.