Diffusion models and the novel Inversion by Direct Iteration (InDI) approach were utilised to translate MRI images into CT images, aiming to avoid harmful CT radiation.
Generative Models, Diffusion Models, Inversion by Direct Iteration, Medical Imaging, MRI, CT, Image-to-Image Translation, Blur
Magnetic resonance imaging (MRI) and computed tomography (CT) are among the most important imaging procedures in medicine. They have complementary strengths as MRI captures soft tissues, especially organs, whereas CT is used when dense structures such as bones should be examined. One of the primary concerns in CT imaging is the use of ionising radiation, which carries potential health risks, particularly for repeated scans and sensitive patient populations. If a translation from MRI to CT was possible, patient exposure to harmful radiation could be avoided.
At the I4DS, a number of projects have already been carried out and several deep learning models have been developed to generate synthetic CT (sCT) images. Previous studies have investigated the use of generative adversarial networks (GANs) for the image translation task. To further advance the field, this bachelor's thesis aims to explore the application of diffusion models for MRI to CT image translation. Diffusion models, a class of generative models, have recently gained attention due to their ability to capture complex data distributions effectively. As a departure from conventional generative model types, the thesis also involves experimenting with a relatively novel approach termed Inversion by Direct Iteration (InDI).
The qualitative evaluation showed that both models generate realistic looking 2D sCT images. Some of them even fooled a domain expert, who mistakenly considered them to be real CT images. In the quantitative evaluation, we assessed competitive performance of our models when compared to the established models by the I4DS. An advantage of DDPMs are their stochastic nature, which enables the visualisation of the prediction's uncertainty within the generated images. However, their main drawback is their iterative and therefore time-intensive image generation process. In contrast, the InDI approach, previously unexplored in the context of image-to-image translation between two modalities, exhibits potential as it required fewer steps than the DDPM and therefore considerably less time to generate sCT images while retaining image quality.
| Type: | Bachelor Thesis (P6) |
| Study Program: | BSc Data Science |
| Duration: | February 21 - August 18, 2023 |
| Team Members: | 2 students |
| Effort: | 720 hours |
Institute for Forensic Medicine
University of Zurich
Winterthurerstrasse 190/Y52
8057 Zürich
irm.uzh.ch
Chantal Strasser (BSc Data Science)
Joseph Weibel (BSc Data Science)
Prof. Dr. Michael Graber (michael.graber@fhnw.ch)
Florian Schärer (florian.schaerer@fhnw.ch)
