Most medical AI you've read about does one thing — spots a lung nodule, flags a stroke, checks a wrist fracture — and charges per scan for that single trick. This week, Alibaba's DAMO Academy broke that mould completely. RADAR, short for Rapid Abdominal Diagnosis with AI and Radiology, is a radiology AI that reads an entire abdominal CT scan and reports on 146 different clinical findings across 18 organs in a single pass. The research landed in Science, one of the world's most selective peer-reviewed journals, on September 18, developed with the First Affiliated Hospital of Zhejiang University.
01.What Alibaba actually built
RADAR is a vision-language model — the same broad family of AI as the chatbots you use daily, but trained to connect medical images with the text in radiology reports. The team trained it on more than 424,000 contrast-enhanced abdominal CT examinations, amounting to 1.5 million image-text pairs and over 15 million anatomy-specific pairs. What's clever is that nobody hand-labelled any of it. The model learned directly from the reports doctors had already written, which is why it scales to so many diseases at once.
Abdominal CT is famously hard for AI. A single scan involves dozens of organs and hundreds of possible conditions, which is why existing tools dodge it and pick narrow single-disease niches instead. RADAR breaks the CT volume into individual anatomical structures and links each one to the matching descriptions in the report — a much more human way of reading a scan.
02.The numbers are genuinely wild
Across nearly 40,000 real-world examinations, RADAR achieved a mean AUC of 0.913 across all 146 findings — for context, 1.0 is perfect, 0.5 is a coin flip, and above 0.9 is considered outstanding in clinical diagnostics. The best competing vision-language model managed 0.776. In emergencies, where it was never specifically trained, it still scored 0.904 across more than 27,000 emergency CT cases. And when tested at eight independent external hospitals, it held an AUC of 0.895 — meaning it generalises beyond its home turf.
Then comes the reader study. Twenty-six specialist radiologists, from junior residents to attending-level physicians, read the same cases independently. RADAR's mean diagnostic accuracy came out ahead of 23 of the 26 doctors. Alibaba is calling it the world's first expert-level generalist medical imaging model.
Here's the thing nobody's talking about though: the more useful number isn't where RADAR beat humans — it's what happened when they worked together. With RADAR as a second reader, the doctors' sensitivity improved by about 10 percentage points, reading time dropped by more than 30%, and junior radiologists matched the accuracy of seniors working unaided. That last bit matters enormously for any country with a radiologist shortage.
03.Why open source is the real headline
Honestly, what surprised me most wasn't the accuracy — it was the price tag. Zero. RADAR's code is up on GitHub under Apache 2.0, and the model weights are downloadable from HuggingFace under a CC BY-NC-SA 4.0 licence, which means free for research and hospital use, with commercial deployment requiring a separate agreement. Any radiology department or research institute can download it, study it, and run it today.
That is a direct shot at the commercial radiology AI industry, which mostly sells narrow single-disease tools on per-scan fees. A free generalist that covers 146 findings in one pass changes the economics for every hospital that has been quoted lakhs for AI subscriptions.
04.What this could mean for Indian hospitals
India runs one of the largest CT scanning workloads in the world, and anyone who has waited three days for a scan report knows the bottleneck isn't the machine — it's the radiologist. A tool that makes juniors read like seniors, and cuts reading time by a third, is exactly the shape of technology our government hospitals and tier-2 diagnostic chains need. It also pairs naturally with the wave of Indian health-AI we've been tracking — when Kardi Ai brought AI-powered long-term ECG monitoring to India, and when Qura Ai's AIRA cleared CE certification, the theme was the same: AI as an assistant that stretches scarce specialists further, not a replacement for them.
05.The catches nobody should ignore
Now the fine print, because there's plenty. RADAR works only on contrast-enhanced abdominal CT — no plain X-rays, no MRI, no brain scans, no non-contrast studies. It has no US FDA clearance and no Indian regulatory approval; any hospital using it diagnostically would be outside clinical boundaries. Its validation was done on Chinese patient populations, so performance on Indian patients would need independent verification. And running it needs real compute and integration with existing PACS workflow. Treat it as a powerful research instrument today, not a plug-and-play product.
For doctors and students watching this space, the practical skill isn't reading the Science paper — it's learning to direct, evaluate, and safely deploy AI tools like this one. Our GPT-6 Astra — Zero to Hero Guide covers exactly that ground: prompting, safe AI agent use, and how to judge what an AI system is actually good at before you trust it.
Key Takeaways
- RADAR reads 146 findings across 18 organs from one abdominal CT scan — trained on 424,000+ exams with no manual labelling
- Mean AUC of 0.913; beat 23 of 26 radiologists; scored 0.904 on emergency cases it was never trained for
- As an assistant, it boosted doctor sensitivity ~10% and cut reading time by over 30%
- Open source: free for research and hospital use (CC BY-NC-SA 4.0 weights, Apache 2.0 code)
- Limits: contrast-enhanced abdominal CT only, no FDA/Indian clearance yet, validated on Chinese patients
Frequently Asked Questions
What is Alibaba's RADAR radiology AI model?
RADAR (Rapid Abdominal Diagnosis with AI and Radiology) is a vision-language model built by Alibaba's DAMO Academy with the First Affiliated Hospital of Zhejiang University. It reads an entire abdominal CT scan and reports on 146 different clinical findings across 18 organs in a single pass. The research was published in the journal Science on September 18.
How accurate is RADAR compared to radiologists?
Across nearly 40,000 real-world examinations, RADAR achieved a mean AUC of 0.913 across all 146 findings, and it outperformed 23 of 26 specialist radiologists in a reader study. With RADAR as a second reader, doctors' sensitivity improved by about 10 percentage points and reading time dropped by more than 30 percent.
Is Alibaba's RADAR model free to use?
Yes, for research and hospital use. RADAR's code is on GitHub under the Apache 2.0 licence and its model weights are downloadable from HuggingFace under a CC BY-NC-SA 4.0 licence. Commercial deployment requires a separate agreement.
What This Means for You
If you're a medical student or radiology resident in India, this is the signal to get fluent with AI now — the juniors who read with AI assistance are already performing like unaided seniors, and that gap will only widen. If you run a diagnostic centre, open-source models like RADAR hint at a future where generalist AI is infrastructure, not an annual licence you negotiate for. And for anyone building in Indian healthtech, the open weights are a research goldmine you can legitimately build on. The fastest way to be ready for this shift is structured learning — the GPT-6 Astra — Zero to Hero Guide is a practical starting point.