Dose-dependent ultrasound neuromodulation of the human brain

TL;DR. The central problem of transcranial ultrasound stimulation (TUS) is that nobody knows how much "dose" actually lands in a given person's brain — skulls and gel coupling vary wildly, and acoustic simulations are shaky on amplitude. This paper measures the dose directly, in vivo, by imaging the micron-scale tissue displacement that ultrasound's radiation force produces (MR-ARFI), and shows that this displacement predicts how strongly the brain responds — across subjects and across voxels within a subject. The simulated intensity everyone currently uses to set dose predicted nothing. If it replicates, this reframes how the field should dose TUS: measure the push, don't trust the model.

Why dosing TUS is genuinely hard

Focused ultrasound can be steered to a deep brain nucleus a transducer can't reach electrically or magnetically, which is why it's the most exciting non-invasive neuromodulation tool right now. But the results are maddeningly inconsistent between people and between studies. The usual suspect is dose: how much acoustic energy actually arrives at the target.

Two things corrupt the dose on its way in. The skull attenuates and refracts the beam in a person-specific way, and the acoustic coupling at the scalp — gel, hair, trapped air — is essentially unmodeled. The standard workaround is to run an individualized acoustic simulation (ray/wave propagation through a CT-like skull model) and report a "simulated in-situ intensity." The problem is that these simulations carry large, acknowledged uncertainty in amplitude. So the field has been setting dose by a number it can't trust, and then wondering why responses scatter.

What you'd want is a ruler that measures the dose where it actually matters: inside the living brain, at the target. That's the gap this paper fills.

The ruler: imaging the push, not the energy

Here is the physical idea. When an ultrasound beam is absorbed and scattered in tissue, it transfers momentum — the acoustic radiation force — which physically shoves the tissue a tiny bit (microns) in the direction the sound is traveling. MR acoustic radiation force imaging (MR-ARFI) makes that shove visible. You turn on a pair of strong motion-encoding gradients in the MR sequence, fire the ultrasound pulse during them, and tissue that moves picks up an extra MR phase proportional to how far it moved:

where \phi is the measured phase shift, \gamma the proton gyromagnetic ratio, G_{\text{MEG}}=70 mT/m the gradient amplitude, and T_{\text{MEG}}=15 ms its duration. The beautiful thing about \Delta x is what it encodes: not the acoustic energy in free water, but the energy as it couples into this tissue's mechanics — absorption and viscoelasticity weighted together. It's the dose the brain actually feels.

The experiment, and why the LGN is the right target

The lateral geniculate nucleus is almost a designed test circuit. It's small, deep, lateralized, and wired monosynaptically to primary visual cortex along a well-mapped retinotopic pathway. So a hemifield checkerboard flickering at 8 Hz drives measurable BOLD in both the LGN and ipsilateral V1, giving you a local readout and a downstream network readout of the same intervention.

In 24 healthy adults, a 64-element phased array (500 kHz, 50 W/cm² free-water I_{\text{SPPA}}, 145 ms continuous pulses) was steered at the left LGN inside the scanner. The design is unusually clean for this field:

• Double-blind, within-subject, counterbalanced. An active control site (temporal grey matter, visually irrelevant) controls for the peripheral confounds — the audible and tactile sensations of TUS — that plague "online" stimulation. Blinding was verified: participants could not discriminate LGN-TUS from control-TUS (d′ = −0.05, not different from zero).
• MR-ARFI measured in each participant (N=21 usable), in an oblique plane through the LGN perpendicular to the beam.
• Simulated in-situ intensity computed with a standard k-Wave pipeline, for a head-to-head comparison against measured displacement.

What they found

First, TUS actually engaged the deep target. LGN-TUS increased task-evoked BOLD in the stimulated LGN relative to control (t(23)=3.12, p=0.005, d_z=0.64), and this increase was specific to the sonicated side — larger than in the contralateral LGN (t(23)=2.13, p=0.044) where there was no effect at all. Prior LGN-TUS work could only infer target engagement from downstream cortex; here it's measured at the nucleus itself. The effect propagated to ipsilateral V1 (cluster-corrected p=0.015), and the V1 effect size was predicted by the LGN effect size (β=0.47, R²=0.22).

Then the headline result. Measured displacement predicted the functional effect — at the LGN (β=0.54, R²=0.29, one-sided p=0.006) and downstream in V1 (β=0.47, R²=0.22, p=0.016). It held within participants too: across voxels in the displacement focus, voxels that got pushed harder showed a larger effect (t(20)=3.41, p=0.003, d_z=0.74).

And the contrast that gives the paper its teeth: simulated intensity predicted nothing — not the functional effect (R²=0.02, p=0.55), and not even the measured displacement itself (R²=0.02, p=0.29).

Crucially, the simulation didn't fail at everything. Its location of peak dose matched the MR-ARFI peak to within ~2 mm — about the registration error of the setup. The divergence was entirely about amplitude, not geometry. That's a precise, useful claim: trust simulations to aim, don't trust them to dose.

A last elegant consistency check. The group-average displacement was biased to the posterior-inferior LGN. By LGN retinotopy, that subregion maps to the upper visual field at central eccentricities — and the TUS-driven V1 cluster landed exactly there (upper-right quadrant, median eccentricity 4.73°, polar angle 75.81°). The downstream effect followed the known wiring, which is hard to explain by any non-specific vascular or auditory artifact.

Why this matters

The live debate in TUS dosing is whether to titrate applied intensity using simulation-derived in-situ estimates to make effects more robust. This paper is direct evidence that doing so on current simulations alone may be chasing a number uncorrelated with reality — the simulated intensity didn't even correlate with the physically measured dose in the same subjects. Displacement, by contrast, tracked the response.

There are two non-exclusive reasons displacement wins. One is practical: simulations don't model hair, air, and coupling, so their amplitudes are a lower bound with large scatter. The other is conceptual and more interesting: intensity is acoustic exposure, while displacement is exposure filtered through the tissue's own absorption and viscoelasticity. If the operative biophysical driver of neuromodulation is mechanical strain on membranes and channels, then displacement is simply closer to the causal variable. MR-ARFI also packages normal strain, shear strain, and particle displacement together — the paper flags that future amplitude/offset manipulations could start to tease these apart in humans.

The most valuable sections to read closely are the "MR-ARFI-quantified dose predicts functional effects" results and the Discussion's treatment of simulation vs. measurement — that's where the generalizable claim lives.

Where to stay skeptical

• Small and single-circuit. N=21 for the displacement analyses, one thalamocortical pathway, fMRI-only readout. The authors explicitly say replication is needed before displacement is used prescriptively. R² around 0.2–0.3 means displacement explains a meaningful but modest slice of variance; most of the between-subject scatter is still unexplained.
• No robust behavior. The contrast-decrement task showed no group effect, blamed on practice effects pushing most subjects off their psychometric threshold (only 8 of 24 stayed in the sensitive 70–80% range) and on the modulated V1 subfield not overlapping the behavioral cue. So this is a physiological result, not yet a perceptual or clinical one.
• One simulation pipeline, one set of acoustic assumptions. The "simulation predicts nothing" claim is as strong as that single k-Wave configuration. A properly benchmarked simulation might do better; the authors say so.
• Dose-response from spatial variation, not dialed intensity. Each participant got a single free-water intensity. The within-subject gradient comes from spatial variation in displacement, not from deliberately varying dose — so the true shape (linear? monotonic? thresholded?) of the dose-response curve is still open. They note rodent work suggests effective displacements an order of magnitude higher, so absolute numbers here are not prescriptive.
• Confounds not fully excluded. The optic radiation runs right next to the LGN and may have been co-stimulated; the mediation test for LGN→V1 propagation had wide bootstrap intervals (indirect effect β=0.18, CI marginally including zero). The retinotopic concordance is the stronger argument for a genuine neural, pathway-following effect.
• MR-ARFI needs an MRI. As a routine dosing tool outside the scanner, it's impractical; its near-term value is as a ground truth to validate and improve the simulations people can run cheaply.

The honest framing, which the authors largely keep, is this: they have shown in humans — for the first time with a verified in-vivo dose metric — that how hard you physically push a deep nucleus predicts how much you modulate its circuit, and that the number the field currently uses to set dose doesn't. That's a real pivot point for how TUS should be calibrated, pending a larger replication and a readout that reaches behavior.