Multi-column network for cell counting

Cell counting is a fundamental but crucial task for microscopic analysis. In this paper, we present a method that can count cells automatically and achieves good accuracy. The algorithm extends the U-net from the single-column to the multi-column to capture the features of cells with various sizes. The general convolutional layers in the U-net body are replaced by residual blocks to help the network converge better. Furthermore, a region-based loss function is designed to guide the model to slide into the proper local minima and avoid overfitting. Experimental results on three public datasets show that the proposed method can handle different kinds of images with promising accuracy. Compared with other state-of-the-art approaches, the proposed approach performs superiorly. (C) 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement