coreMRI: an online web platform for database generation in simulation-based quantitative MR methods

(Abstract presented as a poster at ISMRM workshop on Magnetic Resonance Fingerprinting, October 2017)

Recent studies (Ma et al., 2013 and Xanthis et al., 2015) have shown that the incorporation of Magnetic Resonance Imaging (MRI) simulations could significantly improve quantitative MRI through the utilization of advanced MRI simulations of the pulse sequence on a large population of spins and the generation of a database of simulated MR signals. Despite the increasing interest in incorporating simulation-aided analyses in quantitative MRI, several MR research groups do not have access to a MR simulation platform or the computational resources required to develop an advanced MR simulation platform. In this study we present coreMRI, a MR simulation platform as a web service for database generation in simulation-based quantitative MR methods. coreMRI is publicly available and common across research groups and does not require user expertise on programming or MR physics or upfront investment for purchasing advanced computer systems.

Methods

The simulation framework was developed on the cloud infrastructure provided by Amazon Web Services (AWS - aws.amazon.com) whereas GPU-based instances were utilized through the CUDA-C environment for the computationally intense simulations of MR physics. Distribution and process of data was performed through the MATLAB single-program-multiple data (spmd) framework whereas a dynamic, user-friendly web-page was designed and developed for bridging the user with the MR simulation platform (figure1). The performance and scalability of coreMRI was evaluated on Amazon p2-type instances (1, 8 and 16 GPUs, NVIDIA K80). A T1 mapping pulse sequence for cardiac applications (Messroghli et al., 2004) was simulated on a spin population (5901000 spins) covering a large range of physiological combinations of native myocardial T1 and T2 values. T1 and T2 values of 700–1700 msec and 20–300 msec respectively were simulated with a T1 and T2 step of 1 msec. The excitation slice profile was simulated on 21 spins across the slice-selection direction (Xanthis et al., 2015). The pulse sequence consisted of 157869 discrete time-steps and a database of 281000 entries, each with 8 points, was generated. On a second experiment, a Gradient Echo (GE) pulse sequence was simulated on a spin population (1486905 spins) covering the typical physiological relaxation times of tissues in the brain (Ma et al., 2013). T1 and T2 values of 100–5000 msec and 20–3000 msec respectively were simulated with a T1 and T2 step of 20 msec and 10 msec respectively. For each combination of T1 and T2 values different off-resonance frequencies (B0 inhomogeneity) were simulated covering a BW of 200Hz with an increment of 10Hz. The pulse sequence consisted of 810497 discrete time-steps and a database of 1486905 entries, each with 1000 points, was generated. The second experiment was performed on a p2-type instance with 16 GPUs (NVIDIA K80).

Results

Figure 2 demonstrates the speedup recorded for the spmd framework on the first experiment with 8 and 16 GPUs (NVIDIA K80) compared to a single-GPU configuration. The generation of the database of simulated MR signals took 106 sec for the first experiment and 442 sec for the second experiment on the 16 GPUs computer system.

Discussion

The generation of extended databases of simulated MR signals to be used in simulation-based quantitative MR methods can now be available to the entire MR research community through an online web-platform.

Conclusions

coreMRI (www.coremri.com) is a web service for database generation in simulation-based quantitative MR methods. coreMRI is publicly available and common across MR research groups and does not require programming expertise for the development of an advanced MR simulation platform or upfront investment for purchasing advanced computer systems.

Figure 1. coreMRI web-interface (www.coremri.com) for configuring the characteristics of the database to be used in simulation-based quantitative MR methods.		Figure 2. Speedup of spmd framework with 8 and 16 GPUs (NVIDIA K80) compared to a single-GPU configuration.