Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Generative AI for synthesizing realistic MRI of the developing human brain
The project aims to develop and enhance a novel generative AI-based method for synthesizing MRI images from label map images representing different anatomical brain tissues. Previous works have shown the feasibility of GAN-based methods, but this project focuses on applying stable diffusion networks to synthesize realistic MRI images of developing human brains. The research will also involve evaluating the performance of the proposed method against other approaches, such as GANs and MRI physics modeling-based synthesis, and the resulting code will be made available on GitHub. The project offers an opportunity to contribute to cutting-edge research in the field of fetal and infant brain imaging.
Please note that for patient data confidentiality reasons, we CANNOT ACCEPT CANDIDATES OUTSIDE the UZH/ETH domain or outside the Swiss institutions we collaborate with.
The aim of quantitatively analyzing fetal and infant MRI data is to understand prenatal brain development and establish connections between morphological abnormalities in the developing brain and clinical outcomes or pathologies. This involves segmenting MRIs at various gestational ages into cerebral tissue compartments and anatomical structures.
While manual segmentation is resource-intensive and prone to variability, automated methods offer advantages, and data synthesis approaches could address the bottleneck caused by the limited availability of training data and expert annotations.
Our previous works in the field demonstrated the feasibility of GAN based methods to synthesize MRI.
Stable diffusion networks can conditionally synthesize very realistic looking 2D or 3D datasets, conditional to subject age or pathology. However, they have never been proposed to be used on synthezing MRI of the developing human brain.
In this project, the thesis or semester thesis students will work towards implementing and testing a new generative AI based method that synthesizes MRI from label map images that represent different anatomical tissues of the brain.
The aim of the project is therefore to implement and further improve a stable diffusion network based generative method to synthesize realistic MRI images of development human brains. Second, to evaluate the performance against other methods (GAN, MRI physics modelling based synthesis) and to deposit code on GitHub.
The project will be based at the UZH, group Prof. Andras Jakab, with collaboration provided by the DQBM Group of Prof. Bjoern Menze.
Further reading from our group:
https://www.nature.com/articles/s41597-021-00946-3
https://www.sciencedirect.com/science/article/pii/S1361841523000932
https://arxiv.org/abs/2209.09696
The aim of quantitatively analyzing fetal and infant MRI data is to understand prenatal brain development and establish connections between morphological abnormalities in the developing brain and clinical outcomes or pathologies. This involves segmenting MRIs at various gestational ages into cerebral tissue compartments and anatomical structures. While manual segmentation is resource-intensive and prone to variability, automated methods offer advantages, and data synthesis approaches could address the bottleneck caused by the limited availability of training data and expert annotations. Our previous works in the field demonstrated the feasibility of GAN based methods to synthesize MRI. Stable diffusion networks can conditionally synthesize very realistic looking 2D or 3D datasets, conditional to subject age or pathology. However, they have never been proposed to be used on synthezing MRI of the developing human brain.
In this project, the thesis or semester thesis students will work towards implementing and testing a new generative AI based method that synthesizes MRI from label map images that represent different anatomical tissues of the brain. The aim of the project is therefore to implement and further improve a stable diffusion network based generative method to synthesize realistic MRI images of development human brains. Second, to evaluate the performance against other methods (GAN, MRI physics modelling based synthesis) and to deposit code on GitHub.
The project will be based at the UZH, group Prof. Andras Jakab, with collaboration provided by the DQBM Group of Prof. Bjoern Menze.
Further reading from our group: https://www.nature.com/articles/s41597-021-00946-3 https://www.sciencedirect.com/science/article/pii/S1361841523000932 https://arxiv.org/abs/2209.09696
- to implement and further improve a stable diffusion network based generative method to synthesize realistic MRI images of development human brains
- to evaluate the performance against other methods (GAN, MRI physics modelling based synthesis)
- to deposit code on GitHub, write publication
- to implement and further improve a stable diffusion network based generative method to synthesize realistic MRI images of development human brains - to evaluate the performance against other methods (GAN, MRI physics modelling based synthesis) - to deposit code on GitHub, write publication
