HoVer-Net: Simultaneous Segmentation and Classification of Nuclei in Multi-Tissue Histology Images

Nuclear segmentation and classification within Haematoxylin & Eosin stained\nhistology images is a fundamental prerequisite in the digital pathology\nwork-flow. The development of automated methods for nuclear segmentation and\nclassification enables the quantitative analysis of tens of thousands of nuclei\nwithin a whole-slide pathology image, opening up possibilities of further\nanalysis of large-scale nuclear morphometry. However, automated nuclear\nsegmentation and classification is faced with a major challenge in that there\nare several different types of nuclei, some of them exhibiting large\nintra-class variability such as the tumour cells. Additionally, some of the\nnuclei are often clustered together. To address these challenges, we present a\nnovel convolutional neural network for simultaneous nuclear segmentation and\nclassification that leverages the instance-rich information encoded within the\nvertical and horizontal distances of nuclear pixels to their centres of mass.\nThese distances are then utilised to separate clustered nuclei, resulting in an\naccurate segmentation, particularly in areas with overlapping instances. Then\nfor each segmented instance, the network predicts the type of nucleus via a\ndevoted up-sampling branch. We demonstrate state-of-the-art performance\ncompared to other methods on multiple independent multi-tissue histology image\ndatasets. As part of this work, we introduce a new dataset of Haematoxylin &\nEosin stained colorectal adenocarcinoma image tiles, containing 24,319\nexhaustively annotated nuclei with associated class labels.\n