ROTOR-7 · rotating-gradient MRI

Design concept · not a built machine

Doubling MRI resolution with a gradient field that rotates instead of switches

Spatial resolution in MRI is set almost entirely by how much magnetic field gradient you can impose before the patient’s nerves fire and your amplifiers melt. ROTOR-7 is a concept for getting past both limits at once: rather than three orthogonal coils slamming trapezoidal pulses on and off, a coil array synthesises a single very strong gradient whose direction sweeps continuously around the bore.

The target is 0.25 mm isotropic in‑vivo brain imaging — twice the linear resolution of the best research scanners in operation today.

Read this first. This is a worked engineering proposal, not a product and not a validated result. Nothing here has been built. The rotating-gradient idea sits in a real research lineage (PatLoc, O‑Space, FRONSAC, rotating‑field encoding), but the specific machine described on this page is a design exercise. The section on what it costs and the one on what breaks are the honest parts, and they are where the concept is most likely to fail.

Target voxel

0.25 mm iso

vs 0.5 mm at today’s research frontier

Peak composite gradient

450 mT/m

head insert; clinical whole-body is 40–80

Signal penalty to overcome

 less signal

the voxel is 1/8 the volume — the core problem

Moving parts

0

the field rotates, the hardware does not

01 · Why resolution is stuck

Resolution is a gradient problem, and gradients are a nerve problem

An MRI image is a Fourier reconstruction. Every point you sample lives in k‑space, the spatial-frequency domain, and the finest detail the image can contain is fixed by how far out into k‑space you managed to reach. That reach is the product of gradient amplitude and how long you apply it:

kmax = γ̅ · G · Tread / 2   →   Δx = 1 / (2 kmax) = 1 / (γ̅ · G · Tread) γ̅ = 42.58 MHz/T for protons. To halve Δx you must double the product G · Tread. There is no third option.

So there are exactly two routes to twice the resolution, and one of them is already closed:

Route A — read out for twice as long

Free in hardware terms, and useless in practice. The signal decays with T2*, which at 7 T in grey matter is roughly 25–30 ms and much shorter near air–tissue boundaries. Stretching the readout past that window means the outer, high-frequency part of k‑space is sampled when there is almost no signal left, so you buy blur and distortion rather than detail. Off-resonance error also accumulates linearly with readout time.

Route B — double the gradient

The real lever, and the one every high-performance scanner has pushed on. But raising G the conventional way means switching it faster and harder, and the human body objects: rapidly changing magnetic fields induce electric fields in tissue that depolarise peripheral nerves. Peripheral nerve stimulation (PNS) — not amplifier power — is the binding constraint on modern gradient systems.

This is the crux. PNS does not respond to gradient amplitude. It responds to the induced electric field, which tracks dB/dt — the rate of change — integrated over the spatial extent of the field excursion along the nerve. A conventional Cartesian readout is a train of trapezoids: every ramp is a deliberate, near-maximal dB/dt transient. The scanner spends its entire PNS budget on the ramps, and the ramps contribute nothing to k‑space coverage. They are pure overhead paid in the one currency that is actually scarce.

Where the PNS budget goes in a conventional readout

Trapezoidal gradient waveform against its own rate of change. The shaded spikes are the cost; the flat tops are the only part that images.

The gradient (blue) is useful only on its plateaus. Its derivative (green) — the quantity that stimulates nerves — is zero exactly there, and maximal on every ramp. A sinusoidal drive of the same peak amplitude has a dB/dt that is bounded and spread evenly in time, never spiking. This is the whole basis of the concept.

02 · The concept

Stop switching the gradient. Rotate it.

A conventional scanner has three orthogonal gradient coils (Gx, Gy, Gz) and builds any gradient direction it wants by mixing them. Each coil is driven with a square-ish waveform, and each is sized to survive its own worst case alone.

ROTOR‑7 replaces that with a six-element gradient basis on a head-only insert, driven with mutually phase-shifted sinusoids at a rotation frequency Ω. The superposition of the six is a single linear gradient of constant magnitude whose direction sweeps continuously around the transverse plane:

G(t) = G0 [ cos(Ωt) · + sin(Ωt) · ] Constant magnitude G0, continuously rotating direction. Each of the six coils sees a smooth sinusoid — no coil ever sees a ramp.

Nothing physically spins. “Rotating” describes the field vector, not the hardware — mechanical rotation of a gradient coil at kilohertz rates is not survivable, and it would drag eddy currents and vibration with it. This is the same trick a three-phase induction motor uses to make a rotating field from stationary windings, applied to a gradient set.

Three things this buys you

1. A far higher gradient for the same nerve stimulation. This is the main argument. For a sinusoid, dB/dt peaks at Ω·G0 and is smoothly distributed; for a trapezoid of the same peak amplitude, dB/dt is concentrated into short transients whose instantaneous value is set by the slew rate, typically 5–10× higher. Because PNS thresholds follow a strength–duration (chronaxie) law of roughly 360 µs, brief high-amplitude transients are disproportionately provocative. Removing them entirely lets G0 rise substantially before the stimulation threshold is reached.

2. Thermal load spread across six coils instead of concentrated in one. Each element carries a sinusoid at duty cycle 1/2 rather than a hard-driven square wave. Peak composite gradient is decoupled from any single coil’s continuous rating, which is what normally forces the derating that keeps sustained high-G imaging out of reach.

3. Radial k‑space coverage for free. A rotating gradient traces a radial spoke pattern natively — that is simply what a rotating encoding vector does. You do not pay per-TR rewind and prephase gradients to get there; the trajectory is the drive waveform. Radial sampling oversamples the k‑space centre, which gives self-navigation for motion correction — and at a 0.25 mm target, motion correction is not optional.

The geometry: local, not whole-body

The insert is a 38 cm inner-diameter head-only gradient sitting inside a 60 cm 7 T bore. This matters more than it sounds. PNS threshold scales inversely with the spatial extent over which the induced electric field integrates along a nerve; a compact coil whose field falls away before it reaches the shoulders and torso is intrinsically less provocative than a whole-body gradient at identical amplitude. Head-only inserts are the established route to high gradient strength — the concept here stacks rotation on top of a technique already proven to work.

03 · How it encodes

Watch the encoding vector sweep k‑space

Left: the rotating gradient vector and the six drive currents that synthesise it. Right: the k‑space trajectory it produces. The dashed circle is the k‑space radius a conventional system reaches; everything beyond it is resolution a conventional system cannot obtain.

Gradient field vector

Six coil currents (sinusoids, mutually phase-shifted) summing to one rotating gradient

k‑space trajectory

Radial spokes at the golden angle; dashed ring = conventional kmax

G0 = 450 mT/m  ·  Δx = 0.25 mm  ·  spokes: 0

Drag the gradient slider down to 80 mT/m — a clinical whole-body value — and watch the trajectory collapse inside the dashed ring. That shrinking radius is the resolution loss; the relationship between the two is exactly the 1/kmax of the equation above.

04 · Specification

The machine on paper

Numbers are design targets for the concept, set against a real, strong baseline: a high-performance research head insert, not a clinical scanner. Comparing against a hospital 3 T would flatter the concept and mislead.

ROTOR-7 design targets vs a high-performance research baseline
Parameter Clinical 3 T Research head insert ROTOR-7 (concept) Status
Main field B03 T7 T7 Toff the shelf
Peak gradient G040–80 mT/m200–300 mT/m450 mT/maggressive
Gradient waveformtrapezoidtrapezoidsinusoidthe core change
Rotation frequency Ω1.2 kHzacoustics
Gradient elements336new build
Insert inner diameter60 cm36–42 cm38 cmhead only
Receive channels3232–6496near state of art
Target voxel (isotropic)0.8–1.0 mm0.5 mm0.25 mmunproven
Acquisition time6 min10 min28 minmotion-limited
k-space trajectoryCartesianCartesianradial (golden angle)iterative recon
Anatomy coveredwhole bodyheadheadscope limit

Why 1.2 kHz for the rotation frequency

Ω is a three-way compromise and the least comfortable number in the table. Raising it reduces PNS further (the strength–duration law favours shorter effective pulse widths) and shortens the acquisition. But acoustic output scales steeply with frequency, eddy-current coupling into the cryostat worsens, and 1.2 kHz sits squarely in the band where the human ear is most sensitive. Below roughly 800 Hz the PNS advantage that justifies the whole design starts to erode. If this concept has a single tuning parameter that decides whether it is buildable, it is this one.

05 · The bill

Twice the resolution costs eight times the signal

This is the part of any resolution proposal that decides whether it is real, so it goes before the enthusiasm rather than after it. Signal-to-noise ratio in MRI is proportional to voxel volume and to the square root of total acquisition time:

SNR ∝ Vvoxel · √Tacq   →   halving Δx in 3D ⇒ V ÷ 8 ⇒ SNR ÷ 8 Recovering 8× by scan time alone would need 64× the acquisition — about eleven hours. That is not a scan; that is a disqualification.

So the gradient is necessary but nowhere near sufficient. The rotating coil buys the k‑space reach; something else has to buy back the signal. Here is the full budget, and it is tight:

SNR budget: closing an 8× deficit

Cumulative multiplier against the 8× loss from the smaller voxel. Bars above the line are recovery; the gap at the end is what remains unsolved.

Field strength and coil count are the two honest, well-established gains. Longer scans give a real but sub-linear √T return. Denoising is the entry that deserves the most suspicion — it improves apparent SNR without adding information, and at 0.25 mm the risk of hallucinated structure in a diagnostic image is a genuine safety question, not a cosmetic one.

The honest arithmetic

Multiplying the recovery terms through — 7 T over 3 T gives roughly 2.3×, a 96-channel array over 32 gives about 1.6× in cortex, and 28 minutes over 10 gives √2.8 ≈ 1.67× — lands at 6.15× against a deficit of 8×. The concept is therefore short by roughly 23% in SNR terms before counting any denoising, and it gets there only by comparing a 7 T 28-minute head scan against a 3 T 10-minute one.

Measured against the fairer baseline — the 7 T research insert already achieving 0.5 mm — the field-strength term vanishes and the shortfall is far worse. That comparison is the one that matters, and on it the concept does not currently close. The rotating gradient solves the encoding problem and leaves the signal problem substantially open. Any version of this proposal that does not say so plainly is selling something.

06 · Failure modes

What breaks, in order of how badly

Ranked by the probability of killing the concept outright, not by how hard they are to describe.

Physiological motion at the voxel scale

Critical · may be fatal to the target

The brain is not stationary. Cardiac-cycle pulsation displaces brain tissue by roughly 0.1–0.5 mm, and respiration adds slow drift on top. At a 0.25 mm voxel the target structure moves by one to two voxels within every heartbeat.

This is the risk I would put money on. Radial self-navigation and prospective correction help with bulk head motion, but pulsatile deformation of the tissue itself is not rigid and cannot be corrected by tracking the head. It may simply set a floor on useful in-vivo resolution that no amount of gradient strength can push through.

The SNR shortfall

Critical · quantified above, unresolved

Roughly 23% short against a generous baseline and considerably worse against a fair one. Closing it with learned denoising trades a measurable deficit for an unmeasurable one: the image looks right and you cannot tell from the image whether the fine structure is real. For a diagnostic instrument that trade is not obviously acceptable at any price.

Acoustic output

Serious · engineering-tractable, expensive

A rotating gradient in a 7 T field produces a Lorentz force vector that also rotates, at 1.2 kHz, continuously, with no quiet interval anywhere in the sequence. Conventional scanners are loud in bursts; this would be loud without pause, at the frequency the ear is most sensitive to. Mitigation means vacuum-jacketed formers and force-balanced winding patterns, and it is likely to dominate the mechanical design of the insert.

Eddy currents with no recovery window

Serious · needs active measurement

Conventional sequences have dead time in which induced currents in the cryostat decay. A continuously rotating field never provides one, so the eddy-current field reaches a rotating steady state that distorts the encoding in a way that is systematic rather than transient. Active shielding is necessary but not sufficient; this almost certainly requires concurrent field monitoring with NMR field probes feeding the reconstruction.

Reconstruction cost and non-uniqueness

Moderate · tractable

Non-Cartesian sampling under a time-varying encoding field rules out a plain inverse FFT. Reconstruction becomes an iterative model-based inverse problem that must incorporate the measured field evolution, coil sensitivities, and off-resonance. Computationally heavy, but this is well-trodden ground — it is the least worrying entry here.

PNS modelling confidence

Moderate · the load-bearing assumption

The entire justification rests on the claim that a smoothly rotating field is markedly less stimulating than a switched one at equal amplitude. That follows from the strength–duration law and is well supported for simple waveforms — but PNS thresholds for a continuously rotating field are not well characterised experimentally, and the induced field couples to nerve trajectories in a geometry-dependent way that a rotating vector may excite differently than a linear one. This needs empirical threshold measurement in volunteers before anything else on this page is worth building.

What would settle it, cheaply

The concept is falsifiable well before anyone builds a 450 mT/m insert, and in roughly this order:

  1. PNS threshold measurement on an existing head insert driven with sinusoidal rather than trapezoidal waveforms. Costs weeks, not money, and directly tests the load-bearing assumption. If the advantage is smaller than predicted, stop here.
  2. Motion-floor study: acquire at 0.5 mm with cardiac gating and measure how much of the residual blur is pulsatile deformation rather than noise. This bounds the achievable resolution independent of any hardware.
  3. Six-element field synthesis on the bench with field-probe verification that the composite gradient stays linear through rotation — the coil design is only credible if it does.
  4. Acoustic and eddy-current characterisation at reduced amplitude, before committing to a full-strength build.

Steps 1 and 2 together cost a small fraction of the programme and can kill the idea outright. That ordering is the point: the cheapest experiments test the assumptions most likely to be wrong.