Skip to content

Design — dmipy-design

The acquisition-design layer of the dmipy ecosystem. dmipy-sim simulates a signal from a substrate and a waveform, and dmipy-fit inverts it — but which waveform should you play on a real scanner? dmipy-design answers that: it produces deliverable diffusion gradient waveforms under real hardware limits, and the resulting waveform drops straight into both engines.

Its waveform optimiser is a reimplementation of NOW (Numerical Optimization of gradient Waveforms; Sjölund et al. 2015), re-cast in NumPy/SciPy specifically to integrate with dmipy-sim and dmipy-fit. The deliverability physics and constraints below are established prior work (see References) — not ours; what dmipy-design adds is that ecosystem integration, the min-TE mode, and the Pulseq round-trip.

The newest dmipy engine

pip install dmipy-design
Source: github.com/dmrai-lab/dmipy-design. It shares the same free-waveform interface as dmipy-sim and dmipy-fit, so a designed waveform is a first-class citizen everywhere in the ecosystem.

It works in the instant-pulse approximation (ideal hard RF) — the same regime as the released sim/fit scope — and focuses on the gradient waveform: the thing that actually encodes diffusion. The refocusing-RF designer is a first step beyond that approximation, shaping the 180° pulse itself to stay robust across a real transmit field.

What it does

  • Deliverable waveforms — the gap between idealized dMRI theory (instant, rectangular, symmetric) and what a scanner can actually play (finite slew and amplitude, gradient raster, asymmetric encoding windows, peripheral-nerve-stimulation and heat limits). The NOW design oracle maximises the b-value subject to the full deliverability set, so its output is hardware-realizable by construction — and exports to a scanner-runnable Pulseq .seq.
  • OGSE spectral design — OGSE and PGSE share the same b-tensor; what distinguishes them is spectral content. dmipy-design targets the encoding power spectrum directly, and the designed waveform is simulated in dmipy-sim and fit in dmipy-fit on the same object — one waveform, both engines.
  • Max-b vs min-TE (SNR) — two dual modes: maximise b at a fixed TE, or find the shortest TE that reaches a required b. The latter is SNR-optimal, because a shorter TE means less \(T_2\) decay before the echo.
  • B1-robust refocusing RF — one step beyond the instant-pulse approximation: design the 180° RF envelope to refocus well across real transmit-field inhomogeneity and off-resonance, under peak-B1 and SAR limits — the RF analogue of the NOW gradient box.
  • Run it on the scanner — export a design to a scanner-runnable Pulseq .seq, checked offline (timing, PNS, b-tensor round-trip), so the same waveform you optimise and simulate is the one the scanner plays.

One waveform, three tools

from dmipy_design import design_waveform_now

d = design_waveform_now(b_delta=1.0, G_max=0.08, slew_rate_max=200.0, TE=0.08)  # LTE, Prisma
wf = d.to_sim_waveform()          # -> dmipy-sim: simulate the ground-truth signal
# the same designed scheme drives a dmipy-fit forward model / fit — no conversion layer

The b-tensor shape is set by b_delta (1 = LTE, 0 = STE, −0.5 = PTE); OGSE is a spectral constraint on top; PGSTE reuses the same core with a stimulated-echo timing. Every design is a gradient waveform G(t) — exactly the representation dmipy-sim and dmipy-fit already speak.

References

The constrained gradient-waveform optimisation and its deliverability constraints are prior work; dmipy-design reimplements them in NumPy/SciPy for integration with the dmipy engines and does not claim them as original.

  • NOW — the solver. Sjölund J, Szczepankiewicz F, Nilsson M, Topgaard D, Westin C-F, Knutsson H. Constrained optimization of gradient waveforms for generalized diffusion encoding. Journal of Magnetic Resonance 261 (2015) 157–168. doi:10.1016/j.jmr.2015.10.012. Reference implementation: github.com/jsjol/NOW.
  • Maxwell (concomitant-field) compensation. Szczepankiewicz F, Westin C-F, Nilsson M. Maxwell-compensated design of asymmetric gradient waveforms for tensor-valued diffusion encoding. Magnetic Resonance in Medicine 82 (2019) 1424–1437. doi:10.1002/mrm.27828.
  • Peripheral nerve stimulation — the SAFE model. Hebrank FX, Gebhardt M. SAFE-Model — A new method for predicting peripheral nerve stimulations in MRI. Proc. ISMRM 8 (2000) 2007. Regulatory limit: IEC 60601-2-33 (Consolidated Ed. 3.2, 2015).