Mixed-Supervised Learning for Cell Classification

Cell classification based on histopathology images is crucial for tumor recognition and cancer diagnosis. Using deep learning, classification accuracy is hugely improved. Semi-supervised learning is an advanced deep learning approach that uses both labeled and unlabeled data. However, complex datasets that comprise diverse patterns may drive models towards learning harmful features. Therefore, it is useful to involve human guidance during training. Hence, we propose a mixed-supervised method incorporating semi-supervision and "human-in-the-loop" for cell classification. We design a sample selection mechanism that assigns highly confident unlabeled samples to automatic semi-supervised optimization and unreliable ones for online annotation correction. We use prior human annotations to pretrain the backbone and trustworthy pseudo labels and online human annotations to fine-tune the model for accurate cell classification. Experimental results show that the mixed-supervised model reaches overall accuracies as high as 86.56%, 99.33% and 74.12% on LUSC, BloodCell, and PanNuke datasets, respectively.