Chinese AI Startup Cracks the Century-Old Skull Barrier in Brain Ultrasound

A Chinese hardware and AI startup called Krinwave (鲲为) has obtained medical device registration from China's National Medical Products Administration and deployed its transcranial brain ultrasound system into neurology, neurosurgery, intensive care, and emergency departments at multiple hospitals — doing something the global academic community has treated as largely unsolved for adult patients: imaging the live, blood-flowing brain clearly through an intact skull, without surgery, without contrast agents, and without an acoustic window.

The development arrives at a moment when brain science is generating its own version of the foundation model race. In the past eighteen months, Nature published the first foundation model trained on mouse neural activity, Nature Biomedical Engineering published NeuroSTORM — a general-purpose model trained on 28.65 million fMRI frames from over 50,000 subjects — and Nature Neuroscience published BrainIAC, a brain MRI foundation model trained without labels on nearly 49,000 scans. Each of these efforts immediately ran into the same wall: brain imaging data that is rich enough, continuous enough, and collected at high enough spatiotemporal resolution to train the next generation of neuroscience models does not reliably exist. Krinwave, if its claims are validated independently, is building the infrastructure that could supply it.

Why Transcranial Ultrasound Has Been Stuck for a Century

The physics constraint that Krinwave claims to have cracked is real and well-documented. Sound waves traveling through the skull bone encounter a material whose density and acoustic speed (approximately 3,000 meters per second) differ sharply from both soft tissue (around 1,570 m/s) and the probe's coupling medium. That mismatch produces four compounding problems at once: absorption, refraction, mode conversion — where compressive waves partially convert to shear waves at the bone boundary — and multiple reflections that scatter energy across the image plane. The net effect is that high-frequency sound waves, which carry the short wavelengths needed for fine spatial resolution, are almost entirely absorbed and reflected before reaching brain tissue. Low-frequency waves penetrate better but carry wavelengths so long that small structures in the brain — vessels, deep nuclei, thin cortical layers — fall below the diffraction-limited resolution floor.

The implication that the field internalized over decades is that transcranial fidelity and spatial resolution are mutually exclusive: you can see into the brain, or you can see it clearly, but not both at once. This is why functional ultrasound (fUS) imaging, which maps brain activity by tracking changes in cerebral blood volume with high frame rates, has remained essentially inaccessible in adult human patients without opening the skull. As of the most recent academic literature, functional ultrasound in adults has been applied only during surgery with the skull removed, or in neonates through the fontanel, a soft spot in an infant's skull. Concurrent research on adaptive aberration correction has improved transcranial image quality, but still characterizes the skull as a barrier preventing clinical deployment of fully non-invasive functional ultrasound.

What Krinwave's founder, Jiajia Liu, did between 2009 and 2020 was challenge the foundational assumption that physics dictates the ceiling. Liu holds a background in ultrasound hardware engineering spanning more than a decade of system development before founding the company. The core insight Krinwave describes is that resolution is not a property of frequency alone. It is a property of the system's ability to distinguish sufficiently fine structural, positional, and motion differences from the acoustic data it receives. That ability is, fundamentally, a problem of inference. And inference is something AI systems can be built to do far better than the signal-processing pipelines that conventional ultrasound has used for half a century.

How kOS Turns Low-Frequency Noise Into Clinical Images

The technical architecture Krinwave built to act on that insight is called the HUS (Heterogeneous Ultrasound System) technology system, anchored by a platform the company calls kOS — its proprietary acoustic spatial perception model.

A large language model handles what engineers call a low-bandwidth, high-parameter problem: natural language is a sparse, sequential signal, and models with hundreds of billions of parameters learn statistical relationships across it. Acoustic brain imaging is the inverse: 10 gigabytes of raw signal arrive per second, and the frame rate demand — 5,000 frames per second in Krinwave's system — means there is essentially no time budget for parameter-heavy inference. Building a GPT-scale model for real-time acoustic processing would produce nothing clinically useful; by the time a frame was decoded, the neural hemodynamic event it encoded would have passed.

Krinwave's answer is a small-model architecture the company describes as high-bandwidth and low-parameter — designed specifically for this problem class rather than adapted from existing large model designs. kOS takes large volumes of continuous, heavily degraded acoustic echoes arriving through the skull and converts them in real time into spatial structures and dynamic states. Company materials describe the model's function as analogous to a Transformer for the acoustic domain: it learns the statistical structure of what brain vasculature and its motion look like in the distorted signal space, and inverts that structure to reconstruct clean spatial information.

The imaging pipeline reverses the conventional order. Traditional ultrasound builds images directly from the returning wave pattern — the image is the primary output. Krinwave first builds a three-dimensional physical model of the space: absolute distances, spatial coordinates of structures, blood flow motion vectors. The image a clinician sees is rendered from that model. The company draws an analogy to a game engine: the engine solves the physics of every object's position and collision before it renders the frame. This separation matters because it allows the accuracy of the underlying spatial representation to scale independently of the rendering step.

The compute infrastructure required to run this pipeline in real time is substantial. Krinwave's system operates at what the company describes as over 50 teraflops of floating-point performance alongside more than 1,800 AI TOPS — a compute density achieved by deeply integrating GPU parallelism into the full imaging chain rather than relying on the CPU-plus-FPGA architectures that define conventional ultrasound signal processing. These figures are company-reported and have not been independently verified, but the architectural direction is technically coherent: the step from CPU-based signal processing to GPU-based neural reconstruction follows the same pattern that transformed image recognition, speech synthesis, and a half-dozen other sensory signal problems in the past decade.

The product is built as a five-layer full-stack system. Krinwave designed or specified custom transducer materials at the bottom, dedicated probe geometry for the cranial scenario above that, the high-compute chassis integrating GPU parallel processing in the middle, the kOS model running on that chassis, and clinical diagnosis and treatment software at the top. The company chose full vertical integration rather than adapting existing ultrasound hardware because, it says, the new paradigm cannot run efficiently inside an old hardware architecture — the probe, the compute pipeline, the model, and the clinical interface have to be co-designed to perform correctly at the frame rates and spatial fidelity the system requires.

One consequence of decoupling imaging quality from hardware frequency is that the system's upgrade path changes. Conventional ultrasound advances when transducer materials improve, array densities increase, or analog front-end electronics become more sensitive — hardware generations that arrive over years. A compute-and-model architecture improves when GPU performance increases, when the kOS model architecture is refined, or when training data expands — all of which can occur on a software development cycle. Krinwave reports that its imaging performance improved by more than 100,000 times over five years of internal development. These are company-internal figures and have not been independently audited, but the direction is structurally consistent with how AI-driven signal processing has improved across audio, image, and medical imaging domains since GPU-accelerated deep learning became practical.

The output that matters clinically is a color Doppler imaging system — the type of product NMPA approved, according to the company — that delivers structural brain imaging and real-time blood flow dynamics through the adult skull. Comparison images in Krinwave's materials show a qualitative difference in vessel definition and structural clarity compared with conventional transcranial Doppler. These have not been evaluated in a peer-reviewed clinical study as of this writing. What has occurred, the company states, is regulatory validation: the system passed NMPA's clinical evaluation requirements for a registered medical device and entered hospital use in neurology, neurosurgery, intensive care, and emergency settings.

Where Every Competitor Falls Short

The global race to read brain activity without surgery has attracted some of the best-capitalized teams in technology and neuroscience in 2026. None of them has yet reached what Krinwave claims to have achieved: a registered medical device imaging blood flow dynamics through adult skulls, deployed in clinical departments.

Merge Labs launched in January 2026 with a $252 million seed round anchored by OpenAI and Sam Altman personally, alongside Bain Capital, Gabe Newell, and others. Its technical approach combines functional ultrasound with engineered molecular reporters — specialized proteins inserted into neurons through gene therapy that amplify the mechanical and hemodynamic signatures of neural activity. This is the strategy for overcoming the signal-to-noise problem at the source: make the neural signal louder rather than making the reconstruction smarter. It is a scientifically serious approach, but it requires biological modification of subjects, which places a fundamental barrier between Merge Labs and any near-term clinical deployment in patient populations. Gene therapy for brain-reading purposes requires its own regulatory pathway, long-term safety validation, and patient consent frameworks that do not yet exist at scale. Merge Labs has no registered medical device and no disclosed clinical data as of September 2026. Forest Neurotech, Merge's predecessor research organization, built its early human ultrasound data in patients who had already undergone craniectomy — cases where the skull had been removed for medical reasons, providing a natural acoustic window. That is not the same as non-invasive adult transcranial imaging.

Gestala, a Chengdu-based startup that raised RMB 420 million in its Angel+ round in July 2026, is pursuing focused ultrasound as a neuromodulation tool for chronic pain rather than as a brain imaging modality. The company's clinical pathway targets stimulation — delivering acoustic energy to modulate neural circuits — not the reconstruction of cerebral blood flow maps. Its initial product is stationary clinical hardware; a wearable version is at a conceptual stage. No commercial approval has been sought or obtained, and the company has said it plans to file for Class III medical device registration with the NMPA by the end of 2026 for its first product.

BCI-Sonics, another Chinese focused ultrasound startup, has developed a transcranial phase correction algorithm to address skull-induced distortion of focused ultrasound beams — a problem directly analogous to what Krinwave addresses for imaging, but applied to therapeutic stimulation rather than diagnostic imaging. It is at angel-round stage and targets neuromodulation, not brain imaging.

The established ultrasound industry — Philips, GE HealthCare, Mindray, Siemens Healthineers — does offer transcranial Doppler (TCD) devices for clinical use, but TCD is a fundamentally different modality from what Krinwave is building. Conventional TCD measures blood flow velocity in major cerebral arteries through acoustic windows in the temporal bone and does not produce the spatially resolved, blood-volume-based functional brain maps that make functional ultrasound a neuroimaging modality. TCD sees the broad flow in one vessel; functional ultrasound sees the full vascular tree of a brain region changing second by second.

Academic groups working on transcranial functional ultrasound — including teams at Vanderbilt, the Paris Brain Institute, and research groups at Caltech — are producing the proof-of-concept demonstrations that define what is possible before startups bring it to clinical scale. Their conclusion, as of the most recent literature, is that contrast-free adult transcranial functional ultrasound is achievable but challenging, requiring sophisticated compensation techniques such as compound Barker coded excitation, adaptive motion compensation, and eigen-based clutter filtering — and has not yet been demonstrated in a registered clinical device at scale.

The contrast with Krinwave's claimed position is sharp: the academic literature treats this as an open research problem; Krinwave says it has a registered product deployed in hospitals. That gap cannot be resolved from English-language sources alone and will need to be confirmed through independent clinical data. For now, the company's claims are company-reported, and should be evaluated with that in mind.

Why Brain Foundation Models Make This the Right Moment

The reason the global neuroscience community is paying close attention to the data supply problem is that the scaling pattern that made large language models capable — train on more data, at higher quality, for longer — appears to be producing analogous effects in brain imaging. NeuroSTORM demonstrated that a model trained on 28.65 million fMRI frames from over 50,000 participants generalizes across five downstream tasks including disease diagnosis and demographic prediction, outperforming task-specific supervised models. BrainIAC, trained on nearly 49,000 unlabeled MRI scans through self-supervised learning, matched or exceeded supervised baselines across seven clinical prediction tasks and was particularly effective in low-data scenarios — the authors explicitly noted that the performance gap over supervised models widened when annotated data was scarce, which is the normal state in medicine. The Nature 2025 mouse neural activity foundation model went further: trained on recordings from multiple animals watching natural videos, it generalized to predict responses to entirely new stimulus categories, inferred anatomical cell types, and replicated known connectivity patterns from functional data alone. That result suggests that at sufficient data scale, a single model can begin to capture general principles of brain organization rather than task-specific correlates.

These models expose a data ceiling almost immediately. fMRI collects data in 20-to-30-minute sessions and produces what is accurately characterized as low-frequency static snapshots — the hemodynamic response function, the physiological signal fMRI measures, has an inherent delay of several seconds and a temporal resolution of roughly one to two seconds. That temporal bandwidth is sufficient to capture slow cognitive states and resting-state connectivity networks but inadequate to capture the dynamic, sub-second fluctuations in neural activity that differentiate fine behavioral states from each other. This matters for scaling because a model trying to predict what a specific brain will do next — or decode what it intends — needs temporal resolution that matches the speed of the underlying computation. More fundamentally, brain heterogeneity is extreme: individual brains differ substantially in structure, connectivity, and functional organization, meaning that scaling brain foundation models requires not just more subjects but more data per individual brain — longitudinal, continuous, high-density recordings that capture how a specific brain behaves across many contexts over time. fMRI's infrastructure requirements — 30-to-45-minute sessions in a 1.5- to 3-Tesla magnet costing upward of a million dollars per unit, staffed by trained radiological technicians inside a shielded, acoustically and electromagnetically controlled suite — make continuous longitudinal recording at the volume neuroscience AI needs essentially impractical.

Functional ultrasound addresses each of these constraints on paper. Its temporal resolution runs to sub-100-millisecond frame rates, which approaches EEG-class temporal bandwidth while maintaining the spatial precision of a sectional image. Its sensitivity to cerebral blood volume changes — the same hemodynamic signal fMRI uses, but captured faster and at higher spatial resolution — maps the activity of entire brain regions through a session that can run for hours at a bedside rather than minutes in a scanner suite. Its field of view spans cortex to deep nuclei in a single acquisition, which EEG cannot achieve. A portable device with no superconducting magnet, no RF shielding requirement, and no claustrophobia constraint could in principle be deployed longitudinally in outpatient clinics, ICUs, or eventually in research settings with far lower per-session cost than MRI. If the skull barrier could be crossed consistently in adults without surgery or contrast agents, the result would be a data modality that simply does not currently exist at scale.

Krinwave positions itself precisely here. Jiajia Liu has described the company's role as analogous to Nvidia in the AI hardware ecosystem: not building the top-level foundation models, but supplying the compute and data infrastructure that those models depend on. In practical terms, the company states it provides high-throughput spatial data acquisition, dynamic real-time imaging, and foundational extension capabilities to AI companies, BCI startups, research institutions, and hospitals.

This positioning is strategically careful. It avoids competing with the large laboratory investments building brain foundation models while placing Krinwave at the supply layer that all of those efforts will need. Flourish, backed by Jeff Bezos, Google's GV, Lux Capital, and Catalio Capital at a $500 million raise and $2.5 billion valuation, is pursuing connectomics — studying real neurons at single-cell resolution with electron microscopes to extract the brain's core algorithm — and building silicon implementations of the principles it finds. That work requires reference data about how biological neural circuits actually behave across time and stimulus conditions. OpenAI, investing in Merge Labs, has stated it sees high-bandwidth neural interfaces as upstream infrastructure for AI systems that interpret human intent. None of these long-horizon bets reduce the need for the kind of continuous, non-invasive brain data that only a transcranial functional ultrasound platform can supply without surgical access. Krinwave's commercial restraint — focusing on data acquisition and device deployment rather than attempting to build brain models itself — is the correct call given where the technical bottleneck actually sits.

What Comes Next, and What Remains Uncertain

The ambition Krinwave describes goes beyond imaging. Jiajia Liu has outlined a roadmap toward read-write capability: using focused low-intensity ultrasound to modulate specific brain regions (writing) while simultaneously imaging their response (reading). This is the bidirectional closed loop that the BCI field has long treated as the endpoint of non-invasive neuromodulation — observe what is happening, intervene, and verify the intervention, all without breaking the skin. Independent human studies published in Nature Communications and Neuron in 2025 and 2026 have already shown that low-intensity focused ultrasound can reach deep structures — including the nucleus accumbens and the dorsal anterior cingulate cortex — and produce measurable changes in reward sensitivity and pain perception that replicate across participants. These were strictly research-setting demonstrations using navigated MRI guidance. Krinwave's imaging capability, if it performs as described, would eliminate the need for that MRI guidance step: the imaging system could verify target position and hemodynamic response directly through the ultrasound channel, replacing a workflow that currently requires two modalities with one.

There is also a data governance dimension that international users of Krinwave's devices would need to assess. The company is headquartered in Shenzhen, China. Brain blood flow data collected by its clinical devices falls under China's Data Security Law (2021), which establishes government-access provisions for important data, and China's National Intelligence Law (2017), which requires organizations operating in China to support state intelligence work without exemption for medical device companies. The September 2026 NMPA EEG data quality standard for AI-powered BCI devices — the world's first such binding technical rule — establishes further infrastructure for how neural data from AI-processed BCI devices must be managed. International hospitals evaluating Krinwave hardware should apply the same data governance analysis they would to any medical imaging platform subject to Chinese jurisdiction. This does not affect the scientific question of whether the technology works, but it is a relevant operational consideration that the source material does not address.

The near-term unknowns are significant. No peer-reviewed independent benchmarking of Krinwave's imaging system has appeared in the published literature as of September 2026. The company's performance figures — including the 100,000-fold imaging improvement it describes over five years and the compute specifications it cites — are internal and unaudited. The NMPA registration, while a meaningful regulatory hurdle if confirmed, does not constitute independent scientific validation of the specific performance claims in Krinwave's materials. The clinical institutions reportedly using its devices have not disclosed outcomes data.

These gaps are common for early medical imaging commercialization — peer-reviewed clinical data typically lags regulatory approval — but they do mean the distance between what Krinwave claims and what is independently verifiable remains wide. What is not in doubt is the problem it is solving. The skull barrier has frustrated neuroscientists for decades, the brain foundation model wave has arrived ahead of the data infrastructure needed to sustain it, and no other company in 2026 appears to have crossed the regulatory threshold into hospital-deployed adult transcranial functional brain imaging. The question whether Krinwave's kOS architecture has genuinely dissolved a century-old physics constraint — or compressed it enough for clinical utility under specified conditions — will likely be settled in the peer-reviewed literature in the next twelve to twenty-four months, as hospital deployment data accumulates and independent groups attempt to reproduce its approach. If the data matches the claim, the consequence is not just a better ultrasound machine: it is the first viable route to the continuous, high-resolution, non-invasive brain data that every brain AI project currently lacks.