Generating T2 MRI from T1 Scans Using CycleGAN

CycleGAN pipeline to generate T2-weighted MRI images from T1 scans using unpaired data.

Why I Built This

Paired MRI datasets across contrasts are limited, but multi-contrast information is valuable for diagnosis and downstream models. I built this project to explore whether unpaired image translation could generate useful synthetic T2 scans from T1 inputs. The deeper motivation was to investigate data augmentation approaches that are realistic for constrained medical imaging settings.

Method

  • CycleGAN training on unpaired T1 and T2 MRI domains.
  • U-Net style generators and adversarial training.
  • Combined discriminator, generator, cycle-consistency, and identity losses.

Dataset and Outputs

  • 43 T1 images and 46 T2 images before augmentation.
  • Generated T1->T2 and T2->T1 synthetic outputs, with GIF-based progression visualization across epochs.