DAN : Breast Cancer Classification from High-Resolution Histology Images Using Deep Attention Network

Millions of women succumb to breast cancer every year. Till date, it is mainly diagnosed by core needle biopsy of the breast tissue, followed by analysis of the histopathological image to detect the presence of malignant tumor. In the past few years, deep learning pipelines have been proposed for carcinoma type classification from the breast histology images. They mostly entail in dividing the high-resolution images into patches, followed by classifying the patches using convolutional neural network and finally integrating the patch-wise results for predicting the class of the image. But these methods give the same importance to all the patches and do not focus on the most salient regions of the image. In this paper, we present a novel attention mechanism, which aids the network to specifically focus on the most relevant parts of the image, that is, the design of the network allows for learning a weighted representation of all the constituent patches of an image. Experimental results reveal that our model achieved a $$85.50\%$$ and $$96.25\%$$ for patch- and image-wise classification accuracies, respectively, on the ICIAR 2018 breast histopathological images dataset. Our proposed method outperforms some state-of-the-art methods to the best of our knowledge.