MSA-Net: Multiscale Spatial Attention Network for the Classification of Breast Histology Images

Breast histology images classification is a time- and labor-intensive task due to the complicated structural and textural information contained. Recent deep learning-based methods are less accurate due to the ignorance of the interfering multiscale contextual information in histology images. In this paper, we propose the multiscale spatial attention network (MSA-Net) to deal with these challenges. We first perform adaptive spatial transformation on histology microscopy images at multiple scales using a spatial attention (SA) module to make the model focus on discriminative content. Then we employ a classification network to categorize the transformed images and use the ensemble of the predictions obtained at multiple scales as the classification result. We evaluated our MSA-Net against four state-of-the-art methods on the BACH challenge dataset. Our results show that the proposed MSA-Net achieves a higher accuracy than the rest methods in the five-fold cross validation on training data, and reaches the 2nd place in the online verification.