Abstract The representation of an image as a flow network has gained an increased interest in research for the 2D and 3D segmentation field. One of these segmentation approaches consists in applying a minimum cut algorithm to separate the image in background and foreground. The most remarkable algorithm to segment a 2D image using this approach is GrabCut. This article presents a novel segmentation of 3D image using GrabCut implemented on the GPU. We proposed a scheme where a volume dataset is used as input, instead of a 2D image. The original GrabCut algorithm is adapted to be executed on the GPU efficiently. Our algorithm is fully parallel and is optimized to run on Nvidia CUDA. Tests performed showed excellent results with different volumes, reducing the computation time and maintaining a correct separation background/foreground.