SSalarsNet

AI · Article 7 of 64

Focused Ultrasound and the Possibility of Non-Invasive Deep Brain Control

Can sound waves become a practical way to influence deep neural circuits without surgery?

By Randy Salars · October 7, 2026

Ultrasound is unremarkable technology. A small transducer vibrates, pushing and pulling the water or gel in front of it, and that oscillation travels through tissue as a pressure wave; a machine listens for the echo and draws a picture. Nothing about the procedure suggests a method for reaching the middle of a living human brain and changing what a person does next. Yet the same physics—pressure waves that cross skin, muscle, and bone—underlies one of the more serious candidates for non-invasive deep brain stimulation. The question is whether that candidate can become a practical clinical tool, or whether it will remain a laboratory capability that outruns its evidence.

Focused ultrasound has an unusually clean pitch. Unlike deep brain stimulation (DBS), it needs no implanted electrode. Unlike transcranial magnetic stimulation, it is not confined to the cortex by the geometry of a coil. Unlike transcranial electrical stimulation, it can be aimed. That combination is why the field has drawn funding, media attention, and a steady expansion of human studies. It is also why the claims need careful sorting, because the gap between “we can steer acoustic energy to a deep target” and “we can reliably change a circuit in a way that helps a patient” is where most of the real work sits.

The useful discipline here is to keep three categories separate. There is demonstrated science—effects measured in cells, animals, cadavers, and small human samples under controlled conditions. There is plausible engineering—device systems, targeting models, and safety standards that are reasonable extrapolations from that science but have not yet been validated end to end. And there is speculation—the consumer headsets, the do-it-yourself neuromodulation kits, the idea that deep brain control is a solved problem. Most public discussion collapses these three into one.

Turning a pressure wave into a neural signal

Transcranial ultrasound stimulation (TUS) for neuromodulation typically uses carrier frequencies between roughly 200 kHz and 800 kHz, far below the 1–10 MHz range of diagnostic imaging. The trade-off is physical. Lower frequencies pass through bone with less loss but focus less tightly; higher frequencies focus more sharply but pay for it in attenuation and heating at the skull. A curved transducer or a phased array concentrates the wave into a focal volume, and the intensity at that focus is expressed as spatial-peak pulse-average intensity (I_SPPA). Much of the foundational work in the field used intensities of about 1 to 100 W/cm², delivered in bursts lasting from tens of microseconds to hundreds of milliseconds.

The mechanistic question—how does a gentle mechanical wave make a neuron fire?—has a defensible answer, and it comes from work that is explicitly about cells rather than people. In 2022, Sangjin Yoo and colleagues at Caltech showed that focused ultrasound excites primary murine cortical neurons in culture through a mechanical pathway. The team used 300 kHz ultrasound at 15 W/cm² for 500 ms and recorded calcium responses with a genetically encoded indicator. The neurons responded robustly, with amplitude rising monotonically with intensity and pulse duration, and with an onset delay of roughly 200 ms after stimulation began. The response did not depend on temperature: a fiber-optic thermometer next to the cells recorded a change of 0.005 ± 0.003 °C, and an mCherry temperature reporter stayed flat. It did not require cavitation or large-scale membrane deformation. And it did not require synaptic transmission.

What it did require was specific calcium-selective mechanosensitive ion channels. Pharmacological and genetic inhibition of those channels reduced the ultrasound response; over-expressing them strengthened it. The authors proposed a sequence in which mechanical force opens mechanosensitive channels, calcium accumulates, and calcium- and voltage-gated channels amplify the signal into burst firing. This is a mechanistic explanation for how ultrasound can excite a neuron, and it is the strongest such account available. It is also a result about dissociated cells in a dish, and the authors were careful to note that the acoustic conditions of a culture well are not those of a skull.

The auditory confound that shadows the field

Before the human evidence, one caution deserves to be stated plainly, because it shapes how every in vivo result should be read. Ultrasound delivered through the skull is audible. It excites the cochlea, and the cochlea projects widely.

Two studies published back to back in Neuron in 2018 made this concrete. Guo and colleagues mapped brain activation in guinea pigs and found that ultrasound produced activation across cortex and subcortex—patterns that vanished when the auditory nerve was cut or cochlear fluid was removed. The activation was not primarily a direct effect of sound on neurons; it was the brain hearing the ultrasound. Sato, Shapiro, and Tsao reached the same conclusion with wide-field calcium imaging in mice: ultrasound aimed at visual cortex activated auditory and somatosensory regions, and motor responses looked like a startle reflex that diminished when the animals were chemically deafened.

In humans the confound is measurable too. Braun and colleagues tested whether participants could tell real stimulation from sham and found that they could, using a 500 kHz carrier, a 300 ms burst, and a 1 kHz modulation at 50 percent duty delivered to right visual cortex. Participants showed auditory EEG activation. Critically, playing audio masking reduced detection to chance and abolished the ultrasound-evoked auditory response, and an ex vivo skull showed that bone conducted the pulse-repetition frequency efficiently.

The implication is not that ultrasound cannot affect neurons directly—the calcium work says it can—but that any study that does not control for the auditory pathway risks attributing to a deep target what is actually a hearing-mediated effect. Modern studies increasingly include masking and control targets. That practice is part of what separates a credible result from an evocative one.

What the human studies actually show

Human TUS research has moved from feasibility to targeted circuit effects, and a few 2025 and 2026 papers are worth reading closely. The most cited recent example is a study of the nucleus accumbens by Siti Yaakub and colleagues, published in Nature Communications in 2025. Twenty-six healthy adults received repetitive TUS to the nucleus accumbens in a within-subject design that paired stimulation with functional MRI, used a dorsal anterior cingulate target as an active control, and included sham. The 80-second, 5 Hz protocol changed reward-expectation signals in the nucleus accumbens and shifted behavior: participants adjusted their use of a win-stay strategy, changed how quickly they updated after reward, and repeated rewarded choices differently. The authors noted that DBS of the same structure perturbs similar features, which is a meaningful convergence. They also framed the work as a step toward, not the achievement of, therapy.

A second line of work targets precision rather than a specific behavior. Eleanor Martin and colleagues, reporting in Nature Communications in 2025, described a 256-element helmet array operating at 555 kHz with stereotactic positioning, individualized acoustic planning, and real-time fMRI readout. Aimed at the lateral geniculate nucleus, the system achieved a focal size of about 1.3 mm laterally and 3.4 mm axially—roughly 3 mm³, an order of magnitude smaller than prior deep-brain ultrasound devices. Online stimulation increased visual-cortex activity during a visual task, and a theta-burst protocol decreased it for at least forty minutes afterward. The university press release described this as deep brain stimulation without surgery, which is fair as engineering, but a point from the peer review deserves emphasis: the focal precision was established from acoustic simulations, not from a direct blood-oxygen-level-dependent reading at the target. Modeling is not the same as measurement.

The third example moves toward disease. In 2026, John Eraifej and colleagues reported a randomized, sham-controlled, actively controlled crossover study in Nature Communications involving four men with Parkinson’s disease who already had subthalamic DBS electrodes implanted. Using 130 Hz-pulsed 500 kHz TUS aimed at the internal segment of the globus pallidus, with personalized acoustic lenses and a ventricle as the active control, they found that beta-band power in the ipsilateral subthalamic nucleus fell by about 10.3 percent (95 percent CI 3.8–16.9 percent), that reaction time improved by about 17.7 percent, and that the reduction in subthalamic–motor-cortex beta coupling correlated strongly with the behavioral gain. There was no change in the group’s motor rating scores, and only the low-beta band was reduced where the peaks were separable. This is a proof-of-concept in four people, measuring online effects. It is a genuine result and an important one. It is not evidence that ultrasound treats Parkinson’s disease.

The skull is the hard part

Everything above assumes acoustic energy can reach the target with enough intensity to matter and with enough spatial selectivity to avoid collateral effects. The skull is the obstacle, and the physics is unforgiving.

Attali and colleagues, writing in Brain Stimulation in 2023, built a three-layer model—skin, skull, brain—with absorption and used it to estimate the maximum acoustic transmission coefficient through the human skull for TUS across a range of 100 kHz to 1.5 MHz. Transmission falls as frequency rises and as bone thickens. Across twenty human skulls and twelve beam diameters, the 95th-percentile maximum transmission was 40 to 78 percent—meaning that in the best cases a substantial fraction of the energy survives, and in worse cases much less. The authors also pointed out that the FDA’s standard derating factor of 0.3 dB/cm-MHz, borrowed from diagnostic practice, ignores the insertion loss of the skull entirely and is therefore over-conservative for transcranial work but for the wrong reason—it is not modeling the actual mechanism of loss.

This matters for two reasons. First, it means that a protocol that works in one person may deliver a different dose in another, because skull thickness and geometry vary. Second, it means that the “intensity” written in a methods section is not the intensity at the target; it is the intensity leaving the transducer. Converting one to the other requires a subject-specific acoustic model, which is why the field is moving toward CT-based planning and individualized acoustic lenses. Even then, the model is a prediction through a medium whose properties are heterogeneous, and it is validated in cadavers and phantoms more often than in living brains.

Dosing, safety, and the vocabulary of risk

Because the field has been moving fast without an international standard, a consortium stepped in. The International Transcranial Ultrasonic Stimulation Safety and Standards group (ITRUSST) published a consensus on biophysical safety in Brain Stimulation in 2025. It sets levels for non-significant risk rather than limits that define danger. For mechanical effects, the mechanical index (MI) or the transcranial mechanical index (MItc) should not exceed 1.9. For thermal effects, non-significant risk is met if the peak temperature rise stays under 2 °C, or the absolute temperature stays under 39 °C, or the thermal dose stays under 2 CEM43 in brain, 16 CEM43 in bone, and 21 CEM43 in skin, or if a defined thermal-index threshold for the exposure duration is respected.

The careful language is the point. The authors state explicitly that these are consensus levels, not safety limits, and that exceeding them is not automatically unsafe—there is simply not yet enough data to define where significant risk begins. They also assume participants without contraindications, compromised thermoregulation, vascular fragility, or ultrasound contrast agents. And they note that the consensus informs but does not replace regulatory review or an institutional review board’s judgment about precision of targeting, inclusion criteria, and informed consent.

The practical effect is that the field is being held, by its own practitioners, to conservative limits that predate a full understanding of mechanism. That is appropriate for early human work. It also means that some of the more ambitious proposals—chronic stimulation of deep targets in outpatients, say—currently sit on the plausible-engineering side of the line, because the dosimetry has not been validated for repeated exposure over months.

The same consortium has also attacked a subtler problem: comparability. In 2024, ITRUSST published a consensus on standardized reporting for TUS in Brain Stimulation, specifying what a methods section should disclose—free-field acoustic parameters, the transducer geometry, targeting and neuronavigation approach, and the estimated field at the target. The justification is that a protocol cannot be replicated, meta-analyzed, or safety-reviewed if readers cannot tell what energy reached which structure. This is unglamorous work, but it is the precondition for the field becoming a science rather than a collection of compelling demonstrations. The reporting standard is consensus, not enforcement; whether journals and laboratories actually comply is the test.

Where this could go, and what would have to be true

The honest summary is that ultrasound has demonstrated a mechanism in cells, a plausible set of targeting methods, and a handful of small human effects on circuits and behavior. It has not demonstrated a therapy. Three things would move it.

First, replication with the auditory confound controlled and with targeting verified by something other than the same model used to plan it. Direct measurement at the target—through fMRI readouts, implanted electrodes in patients who already have them, or independent acoustic validation—is the standard that distinguishes a real effect from a plausible one.

Second, dose standardization that travels across skulls. If a protocol cannot specify what energy reached the target in a given participant, it cannot be compared across studies, and it cannot be regulated. Personalized transmission models are the leading candidate, but they need prospective validation.

Third, long-term safety data. The current safety consensus was built for acute exposures. Chronic or repeated use in people without urgent clinical need is a different risk category, and it has not been studied.

Until those exist, the defensible claim is narrow and still remarkable: sound can be aimed at a deep structure, and under careful conditions it can nudge that structure’s activity and the behavior it supports. What that becomes—a surgical alternative for a subset of patients, a research tool for probing circuits, or a consumer product marketed past its evidence—depends less on physics than on how the field handles its own uncertainties.

Sources and further reading

Discussion

What would you add or question? Add your comment below. A human reviews it before publication.

Loading comments…

Join the discussion

Comments are public after approval. Please do not include links, email addresses, or private information. For one short AI reply, address @AIGuide in your comment or reply to its opening comment. Cloudflare verifies submissions to limit spam. Read our community guidelines.

The wider community forum is also open: Browse article discussions in the forum · Forum home