Kunwei’s Transcranial Ultrasound and the Future of Brain Science
In September 2026, brain-computer interfaces (BCIs) once again attracted significant attention as Neuralink expanded its clinical-trial program. The broader BCI industry is moving rapidly, but one fundamental problem remains unresolved: how can the human brain be observed and potentially modulated at high resolution without opening the skull?
Invasive BCIs can place electrodes close to neural tissue and obtain high-quality signals, but the requirement for surgery limits their applicability. Non-invasive technologies such as EEG are much safer and easier to deploy, yet signals must pass through the scalp and skull, reducing spatial resolution and signal quality.
This leaves an important technological gap between the two approaches.
Could ultrasound provide a middle path—penetrating the skull without surgery while delivering sufficiently high spatial and temporal resolution for brain imaging and neuromodulation?
That is the problem Shenzhen-based Kunwei has been pursuing.
Founded in 2019, Kunwei has built its technology around what it calls acoustic spatial intelligence. Its founder, Liu Jiajia, spent more than a decade investigating the theoretical and engineering foundations of high-performance acoustic imaging before applying the approach to transcranial ultrasound.
The company’s central idea is counter-intuitive: instead of continuing to push ultrasound frequency higher to obtain better physical resolution, use lower frequencies for skull penetration and recover image quality through computation.
The approach is computationally expensive—far beyond the workload of conventional ultrasound systems. The rapid growth of GPU computing has therefore become an important enabling factor.
Kunwei’s larger ambition extends beyond selling ultrasound equipment. It is attempting to establish a new infrastructure layer for brain science: a non-invasive, potentially wearable system capable of repeatedly observing the brain and, eventually, combining measurement with stimulation.
🧠 Part 01: The Missing Infrastructure for Brain Science #
In August 2026, the mathematical community saw an unusual example of how neuroscience research can intersect with pure mathematics.
Crouzeix’s conjecture, a long-standing problem involving the field of values of matrices and function-norm inequalities, was reported as solved after more than two decades of research.
One of the researchers associated with this work is Shanmu Jin, a postdoctoral researcher and resident physician in neurosurgery at Peking Union Medical College Hospital.
Jin’s path into matrix analysis was closely connected to another problem: transcranial ultrasound.
The skull presents a formidable obstacle to ultrasound. Bone strongly attenuates, reflects, and distorts acoustic waves, making conventional high-resolution ultrasound imaging of the adult brain extremely difficult.
That limitation exposes a broader problem in neuroscience.
Modern brain research has increasingly sophisticated tools for observing neural activity, but no single technology simultaneously provides all of the following:
- Non-invasive operation
- Whole-brain coverage
- High spatial resolution
- High temporal resolution
- Repeatable measurements
- Potential for continuous or wearable operation
Each existing modality involves trade-offs.
Invasive electrodes can measure electrical activity directly but require surgical access and generally provide localized measurements.
EEG is non-invasive and inexpensive, but scalp and skull attenuation and the inverse nature of the measurement limit spatial resolution.
fMRI provides whole-brain information with excellent spatial coverage, but requires large and expensive systems and is difficult to use for continuous everyday measurements.
Functional ultrasound (fUS) offers extremely high spatiotemporal resolution for measuring cerebral blood-flow dynamics, but conventional systems face the skull barrier.
Functional ultrasound has therefore attracted considerable attention in neuroscience. Research has demonstrated very fine measurements of neurovascular activity, including micrometer-scale observations in animal experiments using advanced localization techniques.
The remaining challenge is obvious: how do you bring that capability through an intact human skull?
Most ultrasound-based brain-imaging approaches have historically avoided the problem by creating an acoustic window or using an implanted interface.
Kunwei has chosen the opposite strategy.
Rather than bypassing the skull, its approach attempts to model and compensate for the skull’s acoustic distortion computationally.
The company describes this as a full-stack acoustic computational imaging system, combining acoustic hardware, signal processing, reconstruction algorithms, and high-performance computing.
That makes the problem fundamentally different from conventional ultrasound engineering.
🔊 Part 02: Why Lower Ultrasound Frequency Could Be the Key #
For decades, ultrasound imaging has operated around a basic physical trade-off.
Higher frequencies provide shorter wavelengths and potentially better spatial resolution. However, they also experience greater attenuation as they propagate through tissue.
Lower frequencies penetrate more effectively but traditionally produce poorer resolution.
This trade-off works well for conventional medical imaging because the imaging target is usually accessible without passing through thick bone.
The skull changes the equation.
High-frequency ultrasound encounters substantial attenuation and distortion at the skull, while low-frequency waves penetrate more effectively but traditionally sacrifice too much resolution to produce useful images.
For years, this effectively made high-resolution transcranial ultrasound an extremely difficult problem.
Kunwei’s approach starts by reversing the conventional assumption.
Instead of asking:
How can ultrasound hardware physically focus low-frequency waves more precisely?
the question becomes:
How much of the missing resolution can be recovered computationally?
This changes the architecture of the entire imaging system.
Computation Replaces Some Physical Constraints #
Traditional ultrasound systems rely heavily on the physical characteristics of the probe and carefully controlled beamforming.
Kunwei’s approach shifts more of that burden into computation.
The basic concept is straightforward:
- Use lower-frequency acoustic waves that can penetrate the skull more effectively.
- Capture the distorted acoustic information with a large-array sensing system.
- Model the propagation and distortion introduced by the skull.
- Reconstruct the image computationally.
- Use GPU acceleration to make the reconstruction fast enough for practical imaging.
The challenge is that computational reconstruction can be extraordinarily expensive.
According to Kunwei’s description, its imaging workload can exceed conventional ultrasound computation by more than three orders of magnitude.
That makes high-performance computing a fundamental component rather than an optional accelerator.
In this sense, Kunwei’s development follows a broader pattern appearing across modern scientific instruments: physical limitations that were once considered fixed can increasingly be traded for computation.
From 0.25 FPS to Thousands of Frames per Second #
Kunwei describes a dramatic increase in reconstruction performance during its development.
Its initial prototype reportedly operated at approximately 0.25 frames per second. The company says current systems can reach thousands of frames per second under its computational imaging pipeline.
The exact performance of a medical imaging system depends on the reconstruction algorithm, acquisition mode, field of view, resolution, hardware configuration, and other parameters, so these figures should not be interpreted as a universal imaging-rate benchmark.
The broader significance is the direction of travel: as GPU computing has become dramatically more capable, computationally intensive acoustic imaging has become increasingly practical.
This creates a feedback loop:
better algorithms → more computation → better reconstruction → more useful imaging → more data → better algorithms.
That is the foundation of the acoustic spatial intelligence approach.
🧬 Part 03: From Ultrasound Equipment to Brain Science Infrastructure #
If transcranial imaging becomes practical, its importance extends beyond a new medical imaging product.
The larger opportunity is brain data.
Modern AI systems have benefited enormously from massive datasets. Brain science faces the opposite problem: high-quality, longitudinal human brain data remains extremely difficult to collect.
The ideal brain-data platform would repeatedly observe the same person across different physiological and behavioral states.
That means measuring the brain during:
- Sleep
- Exercise
- Learning
- Attention
- Stress
- Recovery
- Disease progression
- Therapeutic intervention
Existing brain-imaging systems are poorly suited to continuous, everyday measurement.
This is where wearable ultrasound becomes particularly interesting.
Wearable Brain Imaging #
Ultrasound has an important engineering advantage over CT and MRI: the sensing hardware can potentially be made compact.
Ultrasound arrays can be integrated into head-mounted systems or other wearable form factors, potentially enabling repeated measurements without placing the subject inside a large imaging system.
Research has already demonstrated wearable ultrasound approaches for human brain monitoring, although some experimental systems still rely on surgically created acoustic windows.
Removing that requirement would fundamentally expand the potential user population.
Instead of collecting occasional measurements from relatively small groups of subjects, a non-invasive wearable system could potentially generate longitudinal datasets from large populations.
The resulting dataset would look very different from conventional fMRI collections.
Rather than thousands of isolated scanning sessions, researchers could eventually obtain continuous measurements across months or years.
That could transform the statistical scale of brain research.
From Reading to Read-Write Systems #
Ultrasound also has another unusual property: it can both image tissue and modulate biological activity.
Focused ultrasound has already been investigated for neuromodulation and therapeutic applications.
Combining imaging and stimulation creates a potentially powerful experimental loop:
observe → stimulate → observe again → adjust parameters → repeat.
Instead of treating brain stimulation as an open-loop intervention, researchers could use measurements from the brain itself to guide subsequent stimulation.
Recent experimental work has explored this concept by alternating ultrasound stimulation and functional imaging in animal models.
The long-term vision is a closed-loop brain interface in which the same physical system can measure a neural response, apply an intervention, and immediately evaluate the result.
For neuroscience, this could turn brain experimentation from a sequence of disconnected measurements into an iterative control problem.
🤖 Part 04: Why Brain Data Could Become an AI Infrastructure Layer #
The timing of this technology is closely connected to the broader expansion of AI research.
Large AI laboratories and technology companies are increasingly interested in neuroscience, brain-computer interfaces, and models inspired by biological intelligence.
The reason is straightforward: today’s AI systems still learn from externally generated data, while the brain represents an enormous natural example of an intelligent system operating in the physical world.
But turning that insight into useful AI research requires data.
A model cannot learn much about the brain from occasional snapshots alone.
It needs large-scale, high-quality, longitudinal observations.
This creates an infrastructure problem:
Who will build the equivalent of the data-acquisition layer for brain science?
If non-invasive transcranial ultrasound eventually achieves sufficient resolution, reliability, and portability, it could become one candidate.
The potential combination is unusually attractive:
- Non-invasive acquisition
- Whole-brain observation
- High spatial resolution
- High temporal resolution
- Repeated measurements
- Wearable hardware
- Potential stimulation
- Computational reconstruction
That combination could provide a new foundation for brain-science datasets.
It is important, however, to distinguish the technological possibility from the final outcome. Translating laboratory imaging into a reliable wearable system capable of producing massive, standardized datasets remains a substantial engineering, clinical, regulatory, and scientific challenge.
🧠 Part 05: The Founder Behind Acoustic Spatial Intelligence #
To understand Kunwei’s approach, it is useful to look at its founder, Liu Jiajia.
Liu studied biomedical engineering at Xi’an Jiaotong University and encountered neural networks during his university years.
After entering the ultrasound industry, he focused on a fundamental question:
Can low-frequency ultrasound achieve high-resolution imaging through computation?
At first, this appears to contradict the conventional ultrasound trade-off.
But the underlying reasoning is relatively simple.
If high frequency provides resolution but cannot penetrate the skull effectively, and low frequency provides penetration but sacrifices resolution, perhaps resolution does not need to come entirely from physical wavelength.
Instead, part of the reconstruction problem can be transferred from physics to algorithms.
That shift sounds simple in retrospect.
Turning it into a working medical device is not.
The system requires simultaneous progress in several areas:
- Acoustic propagation modeling
- Array transducer design
- Signal acquisition
- Computational beamforming
- Image reconstruction
- GPU acceleration
- Calibration
- Hardware-software integration
- Clinical validation
Improving any single component is insufficient if the rest of the system cannot keep up.
This explains why the development cycle can span many years.
From Foundational Research to Product #
Liu founded Kunwei in 2019.
The company’s first prototypes followed its early theoretical and engineering work, and by 2024 the technology was being applied directly to transcranial ultrasound.
The company describes this trajectory as a transition from acoustic spatial intelligence theory → imaging technology → medical products → commercialization.
That is different from the more conventional startup path of identifying an existing technology and adapting it to a market.
Here, the starting point was a difficult physical problem.
The company first attempted to establish a new computational approach to acoustic imaging and then searched for applications where that capability could produce a major advantage.
Transcranial brain imaging is arguably one of the most demanding applications because the skull represents an exceptionally difficult acoustic environment.
⚙️ Part 06: Why GPUs Matter to This New Ultrasound Paradigm #
There is a deeper technology trend underneath Kunwei’s story.
For much of the history of medical imaging, performance was constrained primarily by physics and specialized hardware.
Today, computation has become another fundamental dimension of imaging-system design.
A simplified version of the traditional architecture looks like:
better probe → better physical focusing → better image
The computational-imaging architecture looks more like:
better sensing → better models → more compute → better reconstruction
That difference is important.
It means improvements in GPUs, memory bandwidth, parallel algorithms, and AI-assisted reconstruction can directly improve the capabilities of physical instruments.
The same phenomenon has already appeared in computational photography, radar, microscopy, genomics, and scientific simulation.
Ultrasound may be entering the same era.
This is also why Kunwei’s technology is closely associated with GPU computing.
If the company’s reconstruction algorithms genuinely require orders of magnitude more computation than conventional ultrasound, then improvements in accelerator performance can translate directly into imaging performance, latency, resolution, or system cost.
In effect, the imaging system inherits part of the scaling curve of the compute industry.
🌐 Part 07: Could Transcranial Ultrasound Become Brain Science Infrastructure? #
The most ambitious interpretation of Kunwei’s technology is not that it has created another ultrasound machine.
It is that brain measurement itself could become a scalable computing and data-acquisition platform.
The analogy with AI infrastructure is useful, but it should be treated as an aspiration rather than an established market position.
NVIDIA did not become important to AI simply because GPUs were faster than CPUs. Its broader impact came from combining accelerators, software, libraries, developer tools, and an ecosystem into a general-purpose computing platform.
For brain science, a comparable infrastructure layer would require much more than an imaging device.
It would need:
- High-quality transcranial sensing hardware.
- Fast computational reconstruction.
- Standardized acquisition protocols.
- Large-scale longitudinal datasets.
- Clinical validation.
- Software for analysis and experimentation.
- Closed-loop stimulation capabilities.
- A developer and research ecosystem.
If these pieces eventually converge, the resulting platform could support an entirely new generation of neuroscience research.
Researchers could potentially observe the same brain repeatedly, intervene experimentally, collect the resulting data, and use those observations to build increasingly sophisticated computational models.
That would be fundamentally different from today’s fragmented brain-imaging workflow.
🔬 Part 08: The Remaining Scientific and Engineering Questions #
The potential is substantial, but several difficult questions remain open.
First, what spatial and temporal resolution can truly be achieved through an intact human skull?
Laboratory demonstrations and commercial claims need to be evaluated against independently validated measurements, particularly when comparing different imaging modalities.
Second, how generalizable is the technology across individuals?
Human skull thickness, density, geometry, and acoustic properties vary considerably. A system that performs well on one subject may require different calibration or reconstruction strategies for another.
Third, how much computation is required at clinically useful resolution?
A reconstruction pipeline that works in a laboratory can have very different cost and power requirements when deployed in a wearable device.
Fourth, can imaging and neuromodulation be safely combined?
A closed-loop system capable of observing and stimulating the brain introduces additional clinical and regulatory requirements beyond imaging alone.
Finally, can the resulting data become standardized enough for large-scale AI training?
Collecting more data is not sufficient. Brain-science datasets must also be reproducible, calibrated, clinically meaningful, and accompanied by reliable metadata.
These questions will determine whether transcranial ultrasound remains an impressive imaging technology or develops into a broader brain-computing platform.
🚀 Part 09: From Acoustic Intelligence to Brain Science AI #
Kunwei’s story sits at the intersection of three technology curves.
The first is acoustic spatial intelligence: using computational methods to overcome physical limitations in ultrasound imaging.
The second is accelerated computing: GPUs make reconstruction workloads that were previously impractical increasingly feasible.
The third is brain science AI: increasingly capable AI systems create demand for larger, richer, and more continuous datasets describing biological intelligence.
The convergence is potentially powerful.
For decades, the skull has been one of the most stubborn barriers separating brain science from non-invasive high-resolution measurement. Ultrasound has long had the physical ability to penetrate tissue, but conventional imaging architectures could not simultaneously overcome the penetration-resolution trade-off.
Kunwei’s approach attempts to change that equation by moving part of the problem from hardware physics into computation.
If that approach continues to scale, the implications could extend well beyond transcranial imaging.
The ultimate opportunity is a system capable of repeatedly observing, understanding, and eventually influencing the living brain without surgery.
That future is not yet established. It will depend on independent validation, clinical evidence, engineering scalability, safety, and the ability to turn sophisticated reconstruction technology into reliable products.
But the direction is clear: as computation becomes powerful enough to compensate for physical limitations, some problems once considered inaccessible may become computationally tractable.
In that sense, the most important innovation may not be a new ultrasound machine.
It may be a new way of thinking about the relationship between physics, computation, and intelligence—and about whether the brain can eventually become an observable, measurable, and continuously computable system.