Machine Learning Project 2

Abstract

—Image segmentation is a great tool to have insights into the growth of brain neurons in culture. For this, the U-Net neural network architecture can be used to perform neuron and axon segmentation. In this project, we used a modified version of the U-Net which uses a VGG pre-trained network as its encoder. We created two such models which we call UNet11 and UNet16, with the first one using VGG11 and the second one VGG16 as their encoders. We trained the models to detect at first only the neurons cell bodies, and then both neurons cell bodies and axons in microscopic optical images. The neurons were grown either on an empty plate, or on a nanostructured surface, which might influence the direction of the growth of the axons. We thus analyzed the directionality of the axons depending on the orientation of the grid formed by the nanostructure. We show that the models trained to detect the neurons perform well, while the models trained to detect neurons and axons could be improved with a larger dataset of microscopic images with expert labeling of neurons with their axons.