Self-training for Cell Segmentation and Counting

Learning semantic segmentation and object counting often need a large amount of training data while manual labeling is expensive. The goal of this paper is to train such networks on a small set of annotations. We propose an Expectation Maximization(EM)-like self-training method that first trains a model on a small amount of labeled data and adds additional unlabeled data with the model's own predictions as labels. We find that the methods of thresholding used to generate pseudo-labels are critical and that only one of the methods proposed here can slightly improve the model's performance on semantic segmentation. However, we also show that the induced value changes in the prediction map helped to isolate cells that we use in a new counting algorithm.