Segmentation of pelvic tissues via a neural network
Title: Segmentation of pelvic tissues via a neural network
SNIC Project: Berzelius-2022-5
Project Type: LiU Berzelius
Principal Investigator: Alexandr Malusek <>
Affiliation: Linköpings universitet
Duration: 2022-01-25 – 2022-08-01
Classification: 30208


As a part of his master thesis project, Hang Zhao developed (under my supervision) a method for the segmentation of pelvic tissues via the Attention U-Net neural network. The method used a novel approach in obtaining the training images: they were derived from an anthropomorphic XCAT phantom using a CycleGAN neural network. The trained network was then applied to segment images of human patients. Hang did all the work on his home computer with an RTX 1060 Ti GPU having 6 GiB of RAM only. Despite the limitations imposed by the small amount of RAM, the results looked promising. Nevertheless, the resolution of the images had to be downsampled, and the segmentation had to be performed on individual images (2D), not volumetric datasets containing hundreds of images (3D). It is known that segmentation in 3D provides more accurate results. The aims of the proposed project are to (1) evaluate the method on images with a higher resolution, and (2) extend the method to 3D. Results from the first aim would allow us to publish the original method in an impacted journal. If successful, the second aim would result in a novel and powerful method. More information about the prior work can be found in Hang’s thesis “Segmentation and synthesis of pelvic region CT images via neural networks trained on XCAT phantom data”,