Robust and Markerfree Axon Segmentation with CNNs

The automated segmentation of axonal phase-contrast images to allow axonal tracing over time is highly desirable to understand axonal biology in the context of health and disease. While deep learning has become a powerful tool in biomedical image analysis for semantic segmentation tasks, segmentation performance has been limited so far since axons are long and thin objects that are sensitive to under- and/or over-segmentation. We here propose the use of an ensemble-based convolutional neural network (CNN) framework for the segmentation of axons on phase-contrast microscopic images. The mean ResNet-50 ensemble performed better than the max u-net ensemble on the axon segmentation task. We estimated an upper limit for the expected improvement using an oracle-machine. Additionally, we introduced a soft version of the Dice coefficient that describes the visually perceived quality of axon segmentation better than the standard Dice. Importantly, the mean ResNet-50 ensemble reached the performance level of human experts. Taken together, we developed a CNN to robustly segment axons in phase-contrast microscopy that will foster further investigations of axonal biology in health and disease.