The “Open Ground Truth Training Network” (openGTN) research project is about the development of methods for the generation of large quantities of simulated and synthesized magnetic resonance image (MRI) data, with annotations of relevant anatomies. These data are needed for the training and validation of image analysis algorithms (registration, segmentation, quantification), especially for machine learning approaches such as deep learning using convolutional neural networks (goal 1). In openGTN the focus will be on image segmentation.
The project aims to use these simulated MRI data to develop MRI segmentation methods that are as much as possible independent of the specific MR image contrast (T1-weighted, T2-weighthed, etc.), and insensitive to patient and MRI scanner variations (goal 2).
The project also aims to make the simulated MRI data and annotations available in public databases, which can be used for optimizing and benchmarking image analysis algorithms (often called “algorithm challenges”, see e.g. grand-challenge.org) (goal 3).
About the research team
The research in openGTN is performed within a network of 7 R&D institutes spread over Europe. The Medical Image Analysis group (IMAG/e) of the Department of Biomedical Engineering of the Eindhoven University of Technology, The Netherlands, coordinates the project and hosts 3 Early Stage Researchers (ESRs), supported by the highly prestigious and competitive Marie Curie Innovative Training Networks (ITN) fellowship program.
More information about the research team can be found under here.
Research work packages
The openGTN research is organized in 3 work packages (WPs), with 4 specific research objectives (RO).
The simulation of realistic MR image data of the human brain, spine and heart is performed in 2 stages. First realistic ground truth anatomical and functional reference models of these anatomies are created, including variation in shape and tissue properties, and including pathology (Research objective RO1).
Thereafter the JEMRIS (and similar) MRI simulation software is used to generate realistic MR images of these anatomies (RO2) and make these publicly available (RO3). More info about MR simulation with JEMRIS can be found here. Alternatively, we use deep-learning based image synthesis methods to generate MRI data (with so-called generative adversarial networks).
MRI protocol insensitive segmentation
The generated simulated MRI data will be used to optimize segmentation algorithms, especially focusing on making these algorithms as much as possible insensitive to the large variations in image appearances that are usually present in MRI (RO4).
The Marie Curie funding ended officially 31 December 2021. Research is however ongoing, the 3 involved Early-Stage Researchers are still employed at the Eindhoven University of Technology, The Netherlands. They are currently finalizing their scientific journal papers and writing their PD thesis, which will all become publicly available. The many already published conference and journal publications that describe results achieved so far can be found here.
More information, interested to cooperate?
Please contact us via our contact form.