Unsupervised way to segment sperm sample images using CycleGAN for predicting motility and morphology levels

This research is focused on generating sperm segmentations to predict motility and morphology levels of sperms videos in the Visem open dataset. To accomplish this, the CycleGAN architecture will be used with computer-generated synthetic segmented sperm images.

Labeling a segmentation mask of sperm sample images is a highly time-consuming task and it needs a lot of manual human labor. In this master thesis, the Cycle GAN will be researched and used to generate sperm mask images in an unsupervised way. To train cycle gan to transfer real sperm images into segmented images, it is required to use computer-generated synthetic sperm images which look like segmented sperm images.


The goal of this research is to generate appropriate synthetic images which look like segmented sperm images. Then, the researcher will use these computer-generated synthetic images to train a cycleGAN architecture for generating segmented sperm images of real sperm samples images.

Learning outcome

  • Applying a deep generative model for solving segmentation task in an unsupervised way.
  • Getting hands-on experience in implementing and evaluating deep learning models.


  • Python programming
  • Knowledge about deep learning and deep generative models are an advantage


  • Pål Halvorsen
  • Michael Riegler
  • Vajira Thambawita

Collaboration partners

Simula Metropolitan Center For Digital Engineering AS


Visem dataset:

Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

UDCT: Unsupervised data to content transformation with histogram-matching cycle-consistent generative adversarial networks

Deep Multi-Class Segmentation Without Ground-Truth Labels

Conditional Domain Adaptation GANs for Biomedical Image Segmentation

Awesome GAN for Medical Imaging

Contact person