Feature Attention Network for Simultaneous Nuclei Instance Segmentation and Classification in Histology Images

Segmentation and classification of various types of nuclei in tumor tissue histology images is a crucial step in development of computer aided diagnostic systems. Existing techniques for digital profiling of tumor micro environment have common limitations; they require a lot of training data, are computationally costly and don't perform well in challenging scenarios where nuclei exhibit varying inter and intra class characteristics. Hence, to address the challenges of segmenting and classifying nuclei given their vast morphometric properties, we propose a deep learning based model where we use pixel distances from their respective nuclei center points to separate touching and overlapping nuclei. We incorporate attention mechanism to learn complex features of nuclei and refine representation for high accuracy classification. The proposed methodology is assessed on two publicly accessible H&E stained multi-organ histology datasets. We demonstrate higher performance of our model by comparing with recently published algorithms.