Semantic Segmentation of Nuclei from Breast Histopathological Images by Incorporating Attention in U-Net

Breast cancer is a major disease in the world and is detected by histopathological image analysis. The structure and characteristics of nuclei contributes largely in the decision of malignancy of a tumor. There exists several medical image processing techniques based on traditional and CNN methods to segment nuclei from breast histopathological images. However, these algorithms use hand crafted features and depend on availability of large annotated dataset. Moreover, heterogeneous structure and characteristic of nuclei makes it non trivial task. In this context, this paper presents an encoder decoder based CNN architecture to semantically segment nuclei from breast histopathological images. A new attention mechanism is used to extract feature from the nuclei regions at multiple scales. The proposed architecture is evaluated on breast histopathological images and achieved an mIoU of 0.77.