We present in this article a novel methodology for segmenting experimental images of granular suspensions, which uses a convolutional neural network trained on synthetic images generated with a morphological model. In many image processing problems related to physical applications, the lack of annotated data prevents the use of state-of-the-art supervised algorithms. Our solution to overcome this issue is to alleviate the need for annotated images by using a generative morphological model to construct synthetic images subsequently used as training samples. When applied to actual images of a suspension, the convolutional neural network presents good generalization properties and surpasses the performances of traditional segmentation algorithms. This gain in accuracy is crucial to improve the estimation of the local concentration field in the suspension.