ISMRM Raw data format: A proposed standard for MRI raw datasets

Abstract

Purpose: This work proposes the ISMRM Raw Data format as a common MR raw data format, which promotes algorithm and data sharing.

Methods: A file format consisting of a flexible header and tagged frames of k-space data was designed. Application Programming Interfaces were implemented in C/C++, MATLAB, and Python. Converters for Bruker, General Electric, Philips, and Siemens proprietary file formats were implemented in C++. Raw data were collected using magnetic resonance imaging scanners from four vendors, converted to ISMRM Raw Data format, and reconstructed using software implemented in three programming languages (C++, MATLAB, Python).

Results: Images were obtained by reconstructing the raw data from all vendors. The source code, raw data, and images comprising this work are shared online, serving as an example of an image reconstruction project following a paradigm of reproducible research.

Conclusion: The proposed raw data format solves a practical problem for the magnetic resonance imaging community. It may serve as a foundation for reproducible research and collaborations. The ISMRM Raw Data format is a completely open and community-driven format, and the scientific community is invited (including commercial vendors) to participate either as users or developers. Magn Reson Med 77:411-421, 2017. © 2016 Wiley Periodicals, Inc.

Keywords: image reconstruction; magnetic resonance imaging; open source; raw data format.

Figure 1

A minimal ISMRMRD dataset consists of a flexible XML header and raw data organized as sequence of data items consisting of fixed-size data headers and the corresponding k-space data for each set of samples or data chunk.

Figure 2

The raw data structure for each data frame or chunk of the acquisition, consisting of a fixed-size header with encoding numbers, location, etc. and the raw k-space data and (optionally) the k-space trajectory sampling locations. r and i indicate the real and imaginary part of the data points respectively.

Figure 3

The encoding space of a simple 2D Cartesian acquisition. The XML header describes the image encoding and reconstruction fields of view and matrix sizes and k-space sampling bounding box. The image is acquired at a matrix size of 32 × 16 × 1 and field of view of 600mm × 300mm with a slice thickness of 10mm. It should be reconstructed at a matrix size of 16 × 16 × 1 and field of view of 300mm × 300mm with a slice thickness of 10mm. The k-space data were acquired on a grid oversampled by a factor of 2 in the x direction. The “encodingLimits” section indicates the ky center (5) and the minimum and maximum ky values (0 and 11 respectively), i.e. this is a partial Fourier experiment, with a true pixel size that is somewhat larger than the nominal resolution, min(ky) = −5, max(ky) = +6. The example experiment also employs partial Fourier along the readout dimension, i.e. asymmetric echo.

Figure 4

An experimental demonstration where data sets were acquired on scanners from four vendors and converted from the vendor proprietary raw data file formats into ISMRMRD format. A fifth data set was synthesized numerically and stored in ISMRMD format. The five ISMRMRD raw data sets were reconstructed using three image reconstruction programs written in C++, MATLAB, and Python. The resulting images are shown above, from left to right Bruker, General Electric, Philips, Siemens, and the synthetic data set, and from top to bottom C++, MATLAB, and Python reconstruction programs. The differences in SNR and shading between the images are due to the coil geometries: small volume coil vs. knee coil vs. head array.

Figure 5

Images reconstructed with nominal variable density spiral trajectories compared to spiral trajectories stored in ISMRMRD format. ISMRMRD trajectories are predicted using contemporary gradient impulse response functions and therefore produce distortion-corrected reconstructed images.

Figure 6

Image reconstructed from an accelerated EPI sequence using SENSE. The coil sensitivities were estimated from a separate GRE sequence (not shown).