Automated Region Learning Via Cell-Level Labels to Modify Cell Detection Process for Histopathological Images

Recently, cell detection methods based on convolutional neural networks have been extensively studied. However, the distribution of histopathological cells is heterogeneous. Existing methods normally require region-level labeling as guidance to incorporate regional information into the detection process. But the cost of manually regional labeling is higher as the variable shape and the unclear boundary of the region. In fact, different types of histopathological cells with such heterogeneity have certain distribution patterns, meaning that the spatial relationship of cell-level labels also implies regional information. Based on this, we propose a novel cell detection method modified by automated region learning via cell-level labels in histopathological images. We automatically obtain the true region of cell distribution from the cell-level labels based on the Alpha-shape algorithm to train a region-learning network. The results of the region-learning network are used to modify the cell detection process. Extensive experiments indicate the effectiveness of the proposed method.