A Chinese Research Team Just Open-Sourced the Largest Ultrasound AI Dataset Ever — and Medical Imaging AI May Never Be the Same

Artificial intelligence has transformed radiology — for X-rays, CT scans, and MRIs. Ultrasound, the imaging modality used by more people worldwide than any other, has been largely left behind.

The gap is not accidental. The world's largest open-source medical imaging datasets — NIH ChestX-ray14, MIMIC-CXR, CheXpert — all focus on X-ray and CT. These datasets exist at scale because Western hospital systems generate, store, and eventually share chest X-ray data in structured formats that are relatively straightforward to anonymize and annotate. Ultrasound data is different. It is operator-dependent, non-standardized across equipment, and collected in clinical contexts where the imaging quality varies enormously based on who is holding the probe.

The result is that medical imaging AI has been optimized for the imaging tools available to a small fraction of the world's population, while the imaging tool used by the majority — cheap, portable, radiation-free ultrasound — has had almost no open training data at scale.

A research team from Zhejiang University and Hong Kong Baptist University just published what may be the most important open dataset in medical AI this year: US-365K, containing 364,000 ultrasound image-text pairs and 11,676 real clinical cases, accepted at CVPR 2026, and immediately open-sourced on HuggingFace and GitHub. The implications for medical imaging AI extend well beyond academic benchmarks.

What Happened

The paper is titled "Ultrasound-CLIP: Semantic-Aware Contrastive Pre-training for Ultrasound Image-Text Understanding." The authors — Jiayun Jin, Haolong Chai, Xueying Huang, Xiaoqing Guo, Zengwei Zheng, and colleagues from Zhejiang University City College, Hong Kong Baptist University, Zhejiang University, and affiliated clinical institutions — constructed US-365K from hospital records across China.

The dataset contains 364,000 ultrasound image-text pairs paired with clinical report text, covering all major anatomical regions. It also includes 11,676 complete clinical cases. Data validity rate exceeds 90% — unusually high for a clinical dataset at this scale, suggesting careful curation rather than bulk collection.

CVPR 2026, the primary computer vision conference, has an acceptance rate around 25%. Getting a medical dataset paper accepted there signals that reviewers assessed both the dataset quality and the methodological contribution as genuinely significant contributions to the field.

The model that accompanies the dataset — Ultrasound-CLIP — adapts the CLIP (Contrastive Language-Image Pre-training) framework to the ultrasound domain. CLIP, originally developed by OpenAI for natural images, learns to align visual and textual representations by training on matched pairs. Applied to ultrasound, it learns to connect the visual features of a scan with the clinical language used to describe it. Performance on the paper's benchmark: image-to-text retrieval @10 achieves 37.45%, meaning the model returns the correct clinical description in its top 10 results more than a third of the time. Text-to-image retrieval @50 achieves 80.22%.

Both the dataset and the model code are immediately downloadable. No data-sharing agreement. No institutional approval process. No waitlist. Download from HuggingFace in hours.

Why Medical Imaging AI Has Ignored Ultrasound Until Now

The absence of large-scale ultrasound training data is not an oversight — it is the result of properties that make ultrasound data genuinely difficult to work with at scale.

X-ray images look similar across hospitals. The physics of X-ray production is standardized, the DICOM file format is uniform, and chest X-ray reading has been taught the same way for decades. When NIH released ChestX-ray14 in 2017, it was essentially releasing a standardized product. Any model trained on it would generalize reasonably well to chest X-rays from other hospitals because the images share fundamental visual properties.

Ultrasound images do not share that property. The same liver, in the same patient, imaged by two different operators on two different machines on the same day, can produce images that look dramatically different. Probe angle, pressure, frequency, gain settings, patient anatomy, and operator technique all affect the final image. Medical imaging AI that works on ultrasound must account for variance that simply doesn't exist in X-ray or CT.

The annotation challenge is equally severe. Labeling a chest X-ray requires a radiologist. Labeling an ultrasound image requires a specialist who understands real-time operator artifacts, knows how to orient three-dimensional anatomy from a two-dimensional cross-section, and can distinguish genuine pathology from probe-dependent artifact. The expert pool for ultrasound annotation is a fraction of the expert pool for X-ray annotation.

A prior attempt — URFM, the Ultrasound Representation Foundation Model published in 2024 — pre-trained on more than a million ultrasound images using masked image modeling. It could recognize visual patterns in ultrasound. But it lacked paired clinical text, which meant it could not connect what it saw to what a clinician would write. US-365K provides exactly that connection. The 364,000 image-text pairs are the bridge from visual recognition to clinical language understanding — the gap that URFM and every prior approach could not close at scale.

The CLIP adaptation is technically non-trivial. CLIP's original training involved natural images (photographs of objects, scenes, and people) paired with internet-scraped text descriptions. These images have high visual fidelity, clear subjects, and text that describes what a human looking at the image would naturally say. Ultrasound images are grainy, noisy (a phenomenon called speckle artifact), and described in clinical language that assumes substantial anatomical expertise. Teaching a CLIP-style model to align these two domains required genuine methodological innovation — not just scale.

The practical use case makes the technical stakes concrete. A community health worker in rural Mozambique equipped with a $300 handheld ultrasound device (Butterfly iQ and similar products now sell at this price point) and a model trained on US-365K could potentially screen for liver disease, thyroid nodules, or fetal abnormalities — conditions that currently require a specialist radiologist who may not exist within a day's travel. That gap between available imaging equipment and available diagnostic expertise is exactly what AI can close, if the underlying models are both capable and accessible. Open-sourcing US-365K is the prerequisite for that application to exist.

What the Dataset Gets Right — and What It Doesn't

US-365K addresses a real gap with real rigor. But framing it as a solved problem would misrepresent what datasets can and cannot do.

What it gets right is scale. At 364,000 image-text pairs, US-365K is roughly three times the size of MIMIC-CXR, the prior benchmark for open-source clinical imaging with paired text. Scale matters for pre-training: larger datasets produce models that generalize better, fail more gracefully on unusual inputs, and are more robust to the variance that defines real-world clinical deployment. The dataset's coverage of all anatomical regions (rather than focusing on a single organ system, as prior small-scale ultrasound datasets did) is the second critical property — a model pre-trained on US-365K can in principle transfer to any ultrasound application, not just the specific clinical context it was trained on.

The data validity rate above 90% at this scale is notable for a different reason. Many large-scale medical datasets achieve scale by trading off quality — accepting noisy labels, incomplete records, or ambiguous annotations in exchange for volume. A 90%-plus validity rate suggests the team imposed genuine quality thresholds, likely at significant annotation cost. The remaining 10% — roughly 36,000 potentially noisy pairs — is acceptable for pre-training purposes but requires attention when fine-tuning for specific clinical applications where label noise can compound into systematic errors.

The geographic limitation is the most significant caveat. Every clinical case in US-365K comes from Chinese hospitals. This reflects a real and legitimate constraint — the research team had access to Chinese hospital data, and open-sourcing Chinese clinical data at this scale is already an unusually generous decision given the regulatory friction involved in data sharing. But it means the dataset reflects Chinese patient demographics, equipment calibration standards, clinical documentation conventions, and disease prevalence patterns. A model trained on US-365K may perform well for Chinese clinical contexts and require substantial fine-tuning to work reliably in sub-Saharan Africa, South Asia, or Latin America — the populations that most need accessible diagnostic AI.

The retrieval-versus-diagnosis gap is the technical limitation that matters most for anyone considering building on US-365K. The published metrics measure retrieval performance: given an ultrasound image, can the model identify the correct clinical description from a set of candidates? This is a valid and meaningful benchmark. It is not a measurement of diagnostic accuracy — whether the model can generate a clinically correct diagnosis of a new, unseen case. The gap between retrieval performance and diagnostic accuracy is large, well-documented in medical AI literature, and not closed by dataset scale alone. Researchers building clinical applications on top of US-365K will need prospective validation studies that measure the right outcomes.

The honest characterization of US-365K is as enabling infrastructure. ImageNet (2009) did not solve computer vision — it made serious computer vision research possible for anyone with a GPU. US-365K does not solve ultrasound AI. It makes serious ultrasound AI research possible for anyone with a HuggingFace account.

How This Compares to What Exists in Medical Imaging AI

The medical imaging AI dataset landscape has historically been dominated by X-ray and CT data from Western hospital systems. NIH ChestX-ray14 (112,000 X-rays, no paired text) and MIMIC-CXR (227,000 chest X-rays with radiology reports) set the standard for scale in open clinical imaging data. CheXpert (224,000 chest X-rays, Stanford) added uncertainty labeling. All three focus on chest X-ray. None approaches ultrasound at any scale.

The commercial medical imaging AI market for ultrasound includes Caption Health (a GE subsidiary), Butterfly Network, and Sonio — all of which have built diagnostic models on proprietary datasets that are not available to outside researchers. US-365K introduces an open alternative that could enable academic labs and startups without hospital data access to build competitive ultrasound AI for the first time.

Stanford AIMI (AI in Medicine and Imaging) has catalogued US-365K as a recommended resource within days of publication — an institutional endorsement that will accelerate adoption among the research community that uses AIMI's dataset index as a starting point.

The WHO estimates that 2 billion people worldwide have no access to basic diagnostic imaging. Ultrasound, at its current price point and portability, is the most realistic path to expanding imaging access in low-resource settings. The missing piece has consistently been the AI layer — models that could guide non-specialist operators, interpret findings, and flag abnormalities without a radiologist present. US-365K is the first open dataset large enough to plausibly train models capable of that role.

What Comes Next

The early HuggingFace download patterns are the best leading indicator of where US-365K's impact lands first. Researchers in Southeast Asia and Africa have been among the earliest to fork the dataset — consistent with the populations for whom ultrasound AI has the most immediate practical value. The first wave of fine-tuned models targeting specific applications (thyroid screening, cardiac assessment, obstetric abnormalities) should emerge within six months. Researchers and clinicians tracking the evolving landscape of ultrasound AI applications benefit from building a structured second brain to connect dataset releases, model publications, and clinical validation studies into a coherent research picture.

The geographic bias problem is solvable but requires deliberate effort. Research groups in low- and middle-income countries will need to contribute locally annotated data to adapt US-365K-based models to their clinical contexts. This is not a technical barrier — it is a coordination and resource barrier. Partnerships between Zhejiang University and clinical institutions in Africa or South Asia could produce the localization datasets needed to make the global health equity promise real.

The regulatory pathway remains the most significant barrier between US-365K's publication and its clinical deployment in regulated markets. Medical AI datasets face FDA review (US) and CE Mark certification (EU) for any product built on them. Prospective clinical validation — testing a model trained on US-365K against real patients in a prospective study design, measuring sensitivity and specificity against radiologist ground truth — is the required evidence standard. US-365K makes that research possible; it does not make it finished.

The financing dimension matters here. Prospective validation studies for medical AI are expensive — they require patient enrollment, radiologist ground truth annotation, IRB approval, and typically 12-24 months of data collection before publication. Academic labs that download US-365K have the training data but may not have the clinical partnerships or funding to run prospective validation. The gap between "we trained a model on US-365K" and "we have clinical evidence this model is safe for screening use" is filled by grants, hospital partnerships, and commercial relationships with device manufacturers. Butterfly Network, GE HealthCare, and similar companies have the clinical research infrastructure and regulatory experience to bridge that gap — but they are also potential competitors who might prefer the underlying model technology remain proprietary.

The dataset's China-origin also raises a regulatory question for developers targeting Western markets. The FDA's guidance on AI/ML as a Software as a Medical Device (SaMD) requires developers to document training data provenance, including the clinical contexts from which data was collected. A model trained on Chinese clinical data deployed for screening in US hospitals would face questions about whether Chinese disease prevalence, equipment calibration standards, and clinical documentation practices are sufficiently representative of US patient populations. This is not a dealbreaker — it is a requirement for additional validation evidence — but it is a consideration that developers need to plan for.

The longer-term possibility is the most interesting. If models trained on US-365K demonstrate clinical utility — sensitivity and specificity on par with experienced sonographers for specific screening tasks — it creates pressure for healthcare systems globally to invest in portable ultrasound infrastructure. The AI capability unlocks the hardware deployment. The hardware deployment expands access. The expanded access generates new data. The new data improves the models. Whether this cycle begins depends on whether the next six months of research produce results that motivate the investment.

The US-365K release reflects a recurring pattern in AI development: when critical infrastructure becomes open and accessible, the pace of downstream innovation accelerates sharply. ImageNet enabled a generation of computer vision models. MIMIC-CXR enabled chest X-ray AI at scale. US-365K may do the same for ultrasound.

The constraint on that outcome is not technical — the dataset and the methodology are sound. The constraint is whether the global research community has the resources and clinical partnerships to do the fine-tuning, validation, and deployment work that open-sourcing a dataset enables but does not automatically complete.

Keeping track of fast-moving research developments across AI and healthcare is exactly the kind of knowledge management challenge where a structured approach pays off. Tools that help researchers and practitioners capture, connect, and act on AI healthcare dataset developments make it easier to move from reading about a breakthrough to doing something with it.

Frequently Asked Questions

What is US-365K?

US-365K is an open-source ultrasound AI dataset containing 364,000 image-text pairs and 11,676 complete clinical cases, built by researchers from Zhejiang University and Hong Kong Baptist University. It is the largest open-source ultrasound dataset ever created, accepted at CVPR 2026, and immediately available on HuggingFace and GitHub with no access restrictions.

What is Ultrasound-CLIP?

Ultrasound-CLIP is the AI model that accompanies US-365K. It adapts OpenAI's CLIP (Contrastive Language-Image Pre-training) framework to the ultrasound domain, learning to align ultrasound images with the clinical text descriptions from hospital reports. On the paper's benchmark, it achieves 37.45% image-to-text retrieval @10 and 80.22% text-to-image retrieval @50.

How does US-365K compare to existing medical imaging datasets?

At 364,000 image-text pairs, US-365K is roughly three times the size of MIMIC-CXR (227,000 chest X-rays with paired radiology reports), the prior benchmark for open-source clinical imaging with paired text. Unlike prior datasets that focused on a single organ system, US-365K covers all major anatomical regions, enabling models pre-trained on it to transfer to any ultrasound application.

Can US-365K be used to build clinical diagnostic AI right now?

Not directly. The dataset enables pre-training models capable of connecting ultrasound images to clinical language — but the gap between retrieval performance (what US-365K measures) and diagnostic accuracy (what clinical deployment requires) is significant. Products built on US-365K will need prospective clinical validation studies measuring sensitivity and specificity against radiologist ground truth before FDA or CE Mark clearance.

The next landmark for medical imaging AI is not another dataset release. It is the first published prospective validation of a model trained on US-365K, in a real clinical setting, with real patients, measuring outcomes that matter to clinicians. That paper — whenever it arrives — will tell us whether April 2026 was the moment ultrasound AI became serious, or merely another promising start.