MRI systems and ScanHub

A structured way to write pulse sequences

An MRI pulse sequence is a precisely timed program of radio-frequency pulses, gradient waveforms and acquisition windows. On most systems it is written as code against a vendor’s framework, and checking that the hardware can play it out is a separate step.

MR# describes a sequence by its structure. A sequence is assembled by alternating two predefined kinds of time segment: in one kind the gradient amplitudes at the segment boundaries can be chosen freely, in the other they follow from the neighbouring segments. From this description the timing and waveform parameters can be calculated automatically and checked against the limits of the gradient and radio-frequency hardware. A developer works with a small set of objects, such as iterators, elements and parameters, instead of with raw waveforms.

I developed the method with Stefan Röll at Neoscan Solutions, where it was used for a magnetic resonance system for newborns. It is described in a granted patent; the implementation belongs to the company and is not public.

A digital receiver on a graphics card

The MR signal is conventionally demodulated and decimated in dedicated receiver hardware. With Annalena Erbrecht, then a bachelor’s student, and Enrico Pannicke, I examined doing this on an ordinary graphics card instead: the directly sampled signal of a 1.5 T system is demodulated by inverse quadrature modulation and decimated in CUDA.

We compared three decimation filters, a moving average, a cascaded integrator–comb (CIC) filter and a finite impulse response filter, by noise attenuation, processing effort and group delay. The CIC filter gave the best compromise between noise attenuation and effort for this application.

An open platform for acquisition and processing

The console of a scanner usually combines control, reconstruction and storage in one closed system. ScanHub separates them. It is an open-source platform whose functions are divided into separate services, with a web interface on top. Reconstruction and processing can then run on other machines than the one beside the scanner, and can be exchanged without touching the acquisition.

The first demonstration ran a complete MRI workflow in simulation: a pulse sequence was deployed from the web interface, the acquisition was simulated with an open-source Bloch solver, the raw data were reconstructed by a service, and the result was stored and displayed as DICOM images. The project’s current target is open-source low-field MRI hardware.

ScanHub was started at BRAIN-LINK, which I co-founded in 2021, and developed with the Physikalisch-Technische Bundesanstalt in Berlin. I contributed the system and software architecture and am senior author of the abstracts. The implementation is largely the work of David Schote and Johannes Behrens, and the project is now maintained under the scanhub-os (opens in a new tab) organisation.