Real-Time Cell Counting in Unlabeled Microscopy Images

Abstract

Deep learning is largely applied to cell counting in microscopy images. However, most of the existing cell counting models are fully supervised and trained off-line. They adopt the usual training-testing framework, whereas the models are trained in advance to infer numbers of cells in test images. They require large amounts of manually labeled data for training but lack the ability to adapt to newly-collected unlabeled images that are fed to processing systems dynamically. To solve these problems, we propose a novel framework for real-time (RT) cell counting with density maps (DM). It is a semisupervised system which enables training with upcoming unlabeled images and predicting their cell counts simultaneously. It is also flexible enough to allow almost any cell counting model to be embedded within it. With a reliable and automatic training set renewing mechanism, it ensures counting accuracy while optimizing the models by both historical data and new images. To deal with cell variability and image complexity, we propose a Semisupervised Graph-Based Network (SGN) for within the RT counting framework. It leverages a count-sensitive measurement to construct dynamic graphs of DM patches. With the graph constraint, it regularizes an encoder-decoder to represent underlying data structures and gain robustness for cell counting. We have realized SGN along with several baseline networks and state-of-the-art methods within the RT counting framework. Experimental results validate the effectiveness and robustness of SGN. They also demonstrate the feasibility, efficacy and generalizability of the proposed framework for cell counting in unlabeled images.