Category Separation for Weakly Supervised Multi-Class Cell Counting

Cell quantification in immunohistochemically stained images assists the prognostic evaluation of tumor progression. The development of lighter artificial intelligence-assisted diagnosis models can further improve the efficiency of whole slide image analysis and control training costs. However, recent count-level regression models are limited in the multi-category cell counting task. Therefore, we propose a cellular quantification framework under the count-number supervision by introducing a category-separation learning method, which instructs the model to discriminate differences between various cells. Moreover, a novel loss function is presented to control the representation learning of multi-category cells. In addition, we also validate the performance of Transformer-based architecture. By comparing with learning methods under dots-labeled supervision, the proposed method realize the simultaneous counting of multi-category cells and exhibit competitive results on the public dataset.