Abstract Researchers address the generalization problem of deep image processing networks mainly through extensive use of data augmentation techniques such as random flips, rotations, and deformations. A data augmentation technique called mixup, which constructs virtual training samples from convex combinations of inputs, was recently proposed for deep classification networks. The algorithm contributed to increased performance on classification in a variety of datasets, but so far has not been evaluated for image segmentation tasks. In this paper, we tested whether the mixup algorithm can improve the generalization performance of deep segmentation networks for medical image data. We trained a standard U-net architecture to segment the prostate in 100 T2-weighted 3D magnetic resonance images from prostate cancer patients, and compared the results with and without mixup in terms of Dice similarity coefficient and mean surface distance from a reference segmentation made by an experienced radiologist. Our results suggest that mixup offers a statistically significant boost in performance compared to non-mixup training, leading to up to 1.9% increase in Dice and a 10.9% decrease in surface distance. The mixup algorithm may thus offer an important aid for medical image segmentation applications, which are typically limited by severe data scarcity.