Abstract
Advances in microscopy have increased output data volumes, and powerful image analysis methods are required to match. In particular, finding and characterizing nuclei from microscopy images, a core cytometry task, remains difficult to automate. While deep learning models have given encouraging results on this problem, the most powerful approaches have not yet been tested for attacking it. Here, we review and evaluate state-of-the-art very deep convolutional neural network architectures and training strategies for segmenting nuclei from brightfield cell images. We tested U-Net as a baseline model; considered U-Net++, Tiramisu, and DeepLabv3+ as latest instances of advanced families of segmentation models; and propose PPU-Net, a novel light-weight alternative. The deeper architectures outperformed standard U-Net and results from previous studies on the challenging brightfield images, with balanced pixel-wise accuracies of up to 86%. PPU-Net achieved this performance with 20-fold fewer parameters than the comparably accurate methods. All models perform better on larger nuclei and in sparser images. We further confirmed that in the absence of plentiful training data, augmentation and pretraining on other data improve performance. In particular, using only 16 images with data augmentation is enough to achieve a pixel-wise F1 score that is within 5% of the one achieved with a full data set for all models. The remaining segmentation errors are mainly due to missed nuclei in dense regions, overlapping cells, and imaging artifacts, indicating the major outstanding challenges.