Comparing the segmentation of quantitative phase images of neurons using convolutional neural networks trained on simulated and augmented imagery.

Significance: Quantitative phase imaging (QPI) can visualize cellular morphology and measure dry mass. Automated segmentation of QPI imagery is desirable for tracking neuron growth. Convolutional neural networks (CNNs) have provided state-of-the-art results for image segmentation. Improving the amount and robustness of training data is often crucial to improving CNN output on novel samples, but acquiring enough labeled data can be labor intensive. Data augmentation and simulation can be used to address this, but it is unclear whether low-complexity data can result in useful network generalization. Aim: We trained CNNs on abstract images of neurons and on augmented images of real neurons. We then benchmarked the resulting models against human labeling. Approach: We used a stochastic simulation of neuron growth to guide abstract QPI image and label generation. We then tested the segmentation performance of networks trained on augmented data and networks trained on simulated data against manual labeling established via consensus of three human labelers. Results: We show that training on augmented real data resulted in a model that achieved the best Dice coefficients in our group of CNNs. The largest percent difference in dry mass estimation with respect to the ground truth was driven by segmentation errors of cell debris and phase noise. The error in dry mass when considering the cell body alone was similar between the CNNs. Neurite pixels only accounted for Conclusions: Augmented data outperformed the simulated abstract data for this testing set. The quality of segmentation of neurites was the key difference in performance between the models. Notably, even humans performed poorly when segmenting neurites. Further work is needed to improve the segmentation quality of neurites.