Abstract
The visual attributes of cells, such as the nuclear morphology and chromatinopenness, are critical for histopathology image analysis. By learningcell-level visual representation, we can obtain a rich mix of features that arehighly reusable for various tasks, such as cell-level classification, nucleisegmentation, and cell counting. In this paper, we propose a unified generativeadversarial networks architecture with a new formulation of loss to performrobust cell-level visual representation learning in an unsupervised setting.Our model is not only label-free and easily trained but also capable ofcell-level unsupervised classification with interpretable visualization, whichachieves promising results in the unsupervised classification of bone marrowcellular components. Based on the proposed cell-level visual representationlearning, we further develop a pipeline that exploits the varieties of cellularelements to perform histopathology image classification, the advantages ofwhich are demonstrated on bone marrow datasets.