1 – MRI generation approaches
The original approach of openGTN was to generate MR image data by means of Bloch simulators like JEMRIS (section 2 below). However, in the first project year it became clear that the novel technique of deep learning with neural networks, especially with so called generative adverserial networks (GANs), can also very well generate very realistic MR image data (section 3 below). We therefore decided to investigate and compare both approaches.
2 – Simulated MRI with JEMRIS
We use the JEMRIS simulator as starting point in the openGTN project to generate realistic simulated MRI of the brain, spine and heart, including anatomy, function and pathology:
Detailed information about JEMRIS can be found in this publication:
T. Stoecker et al., High-performance computing MRI simulations, Magn Reson Med 64:186–193 (2010)
and the software can be downloaded from the JEMRIS website.
Details on the application of JEMRIS in the openGTN project can be found here. The European Society of Magnetic Resonance in Medicine and Biology (ESMRMB) organizes a course on the use of JEMRIS once every 2 years.
Example application
An example of the successful application of JEMRIS to generate realistic simulated MRI data of the carotid arteries can be found in this publication:
Harm A. Nieuwstadt et al., Numerical Simulations of Carotid MRI Quantify the Accuracy in Measuring Atherosclerotic Plaque Components In Vivo, Magn Reson Med 72:188–201 (2014).
In this paper, we describe how we successfully used JEMRIS to generate very realistic simulated carotid MR images with ground truth plaque segmentations, and used these to assess the accuracy of manual plaque component segmentation. The included plaque components were: lipid-rich necrotic core (LRNC) and fibrous cap (FC).
The figure below show a comparison between real and simulated MRI for the same plaque geometry and composition. It can be seen that the intensity patterns along the line A, B, C, D, E are very similar.
3 – Data synthesis with GANs
The recently presented paper at the Medical Imaging with Deep Learning conference (MIDL 2020) reports our first results of GAN-based cardiac MR image synthesis (ArXiv).
4 – Public database
We offer our simulated and synthesized data for non-commercial use, more information can be found at opengtn.eu/database.