NORMA eResearch @NCI Library

Generating MRI images using style transfer learning

Mahajan, Dnyaneshwari Sudhir (2023) Generating MRI images using style transfer learning. Masters thesis, Dublin, National College of Ireland.

[thumbnail of Master of Science]
PDF (Master of Science)
Download (5MB) | Preview
[thumbnail of Configuration manual]
PDF (Configuration manual)
Download (10MB) | Preview


In the medical field, the availability of medical imaging for practitioners or radiologists is a huge challenge because of the cost of medical imaging and patient data privacy laws. Deep learning has emerged as an important technology in computer vision applications due to technologies like GAN compared to traditional augmentation techniques. State-of-art style transfer (Generative Adversarial Networks) GANs play a crucial role in generating images in the medical field. The aim of this research project is to build a model for generating MRI (Magnetic resonance imaging) images with different contrast levels using a novel approach of combing the CycleGAN framework and customized U-Net segmentation(CycleGAN+UNET).

This research conducted four experiments on 2 publicly available unpaired datasets of T1-styled images and T2-styled images. At the training phase, the Cyclic consistency loss for the final model is 3.75 which is comparatively low than other models. The generator loss and discriminator loss are 0.93 and 0.79 respectively, the values of the generator and discriminator are very close to each other which means they are balanced. The identity loss is 0.20 which is comparatively low than other models and indicates that the reconstructed image is similar to the original image. The result shows that model performance is better when generating samestyled images T1 to T1 and T2 to T2 with maximum SSIM (Structural similarity) scores of 0.87, 0.94 and PSNR (Peak-signal-to-noise-ratio) scores of 33.09db, and 30.72db respectively. Improvement is needed to generating images with different styles images T1 to T2 and T2 to T1 because of low SSIM scores of 0.35 and 0.55 respectively.

Item Type: Thesis (Masters)
Ul Ain, Qurrat
Uncontrolled Keywords: GAN; CycleGAN; MRI; SSIM; PSNR; U-Net; Deep learning
Subjects: Q Science > QA Mathematics > Electronic computers. Computer science
T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science
T Technology > Biomedical engineering
Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions: School of Computing > Master of Science in Data Analytics
Depositing User: Tamara Malone
Date Deposited: 19 May 2023 16:00
Last Modified: 19 May 2023 16:00

Actions (login required)

View Item View Item