An improved U-Net for cell confluence estimation

Abstract

Cell confluence is an important metric to determine the growth and the best harvest time of adherent cells. At present, the evaluation of cell confluence mainly relies on experienced labor, and thus it is not conducive to the automated cell culture. In this paper, we proposed an improved U-Net algorithm (called DU-Net) for the segmentation of adherent cells. First, the general convolution was replaced by the dilated convolution to expand the receptive fields for feature extraction. Then, the convolutional layers were combined with the batch normalization layers to reduce the dependence of the network on initialization. As a result, the segmentation accuracy and Fl-score of the proposed DU-Net for adherent cells with low confluence (<50%) reached 96.94% and 93.87%, respectively, and for those with high confluence (≥50%), they reached 98.63% and 98.98%, respectively. Further, the paired t-test results showed that the proposed DU-Net was statistically superior to the traditional U-Net algorithm.