The efficiency of many chemical engineering applications depends on the surface/volume ratio of the dispersed phase. Knowledge of this particle size distribution is a key factor for better process control. The challenge of measurements acquired by optical imaging techniques is the segmentation of overlapping particles, especially in high phase fraction flows. In this work, a convolutional neural network is trained to segment droplets in images acquired by a shadowgraphic approach. The network is trained on artificial images and implemented into a droplet size algorithm. The results are compared to an OpenSource segmentation approach.