Introduction

Quantitative MRI, i.e. measuring biophysical parameters using MRI, is a very active research field and can be traced back to the early days of NMR with Hahn's method to measure T₂¹. In this abstract, we describe a new Python-based toolbox, also providing command-line tools, a dependency handler to easily define complex pipelines, and access to imaging meta-data, including vendor-specific fields.

Design and features

To balance the ease of use with the potentially complex qMRI pipelines, Erwin has two main interfaces: a set of command line tools and a Python API. For quick tests or for the simpler methods, the command-line tools hide the details of the Python API in a unified syntax, reachable by the main executable "erwin": a list of all methods can be obtained by calling "erwin --list", the details of each method and the corresponding arguments are given by calling e.g. "erwin b1_map.afi --help". These command-line tools are automatically defined from the Python API, which ensures feature parity between the two interfaces. The Python API will prove useful for more complex pipelines, e.g. the T₂ mapping based on bSSFP² which requires, in addition to a set of bSSFP images with varying flip angles and phase increments, a B₁-map and a T₁-map. This is accomplished in Erwin by defining the inputs and outputs of each method, and by connecting the output of a task to the input of the next one, as shown in Figure 1; in this example, the T₁-map is computed using the VFA method³, which requires the B₁ map. This effectively builds a directed dependency graph between steps, with the property that changing one of the steps does not require recomputing the upstream steps, only the downstream ones: in the previous example, replacing the T₁-map will only recompute the T₂-map, while replacing the B₁-map, also used in the T₁-map, will recompute first the T₁-map, then the T₂-map, ensuring the result is always up to date, with a minimal computation cost. The pipeline engine in Erwin also allows to generate a pictorial representation of a task graph, shown in Figure 2 for the example of the T₂-map pipeline.
One common characteristic of many qMRI methods is their use of acquisition parameters, e.g. flip angle, repetition time or b-value. These parameters are usually stored either in the DICOM files produced by the scanner or in separate meta-data files such as the BIDS sidecar files. In the case of DICOM, some of the parameters are included only in vendor-dependent fields, as is the case for the phase increment of the T₂-bSSFP method cited above. Moreover, not all formats use the same unit for a given field: DICOM store the timing parameters (echo time, repetition time, inversion time, etc.) in milliseconds, while BIDS uses seconds. In order to be vendor-independent, Erwin provides tools to access meta-data stored in either DICOM or JSON format, with the option to convert it from its specific unit. For example, the inversion times of a pulsed ASL sequence⁴, stored in the Siemens private field using microseconds can be accessed by "erwin meta_data.get -p XXX.dcm -q 00291020.MrPhoenixProtocol.0.alTI".
Erwin does not strive to re-implement all qMRI methods: when a complete and well-known toolbox exists for a specific set of methods, e.g. MRtrix⁵ for diffusion or MEDI⁶ for QSM, the provided tools are simply wrapped as tasks so that they can transparently be included in pipelines. This is straightforward for pre-compiled executables or Python libraries, and, for MATLAB-based code, we provide a bridge between the MATLAB engine and the Python runtime.

Conclusion

Erwin has been released under free software license (MIT), is available on GitHub⁷ and PyPi⁸, and includes a documentation of all provided methods⁹, and a set of reference images with their meta-data¹⁰. The methods included as of the writing of this abstract target field mapping (B₀ and B₁), relaxometry (T₁, T₂ and T₂*), cerebral blood flow, diffusion, magnetization transfer, and susceptibility. Although obviously not complete due to the ever-growing number of publications regarding qMRI, the features already present in Erwin and its open development model can benefit the whole qMRI community.

Acknowledgements

No acknowledgement found.