What Happened
On September 18, Alibaba's DAMO Academy published a paper in Science describing a system that reads a single contrast-enhanced abdominal CT scan and reports back on 146 separate clinical findings spread across 18 organs. Liver cancer, pancreatic tumors, gastric and colorectal malignancies, fatty liver, acute appendicitis — one pass, one model. Then the team did the part that almost nobody at this level of capability does in healthcare: they posted the code on GitHub, the weights on Hugging Face, and let anyone download the whole thing.
The model is called DAMO RADAR, and its training data explains a lot about why it exists at all. The team gathered 420,000 contrast-enhanced abdominal CT examinations, each already paired with the radiology report a working doctor had typed up at the time. Those reports became roughly 15 million anatomy-aware image-text pairs. No one sat down and hand-labeled lesions. The supervision was already sitting in hospital archives, written by radiologists doing their jobs, and the training pipeline simply learned to read it.
The headline numbers are unusually specific for a field that tends to publish press releases rather than evidence. Across nearly 40,000 real-world examinations validated at eight independent clinical centers, RADAR posted a mean area under the curve of 0.913 over all 146 findings. AUC measures how well a test separates the sick from the healthy: 0.5 is a coin flip, 1.0 is perfect, and anything above 0.9 is conventionally described as outstanding. In a formal reader study against 26 specialist radiologists drawn from several hospitals — a mix of junior residents and attending physicians — RADAR's mean diagnostic accuracy came out ahead of 23 of them.
Getting there required solving a problem that trips up standard vision-language models. A CT volume is three-dimensional and overwhelmingly boring: the vast majority of voxels are ordinary tissue, and whatever is wrong tends to be small and localized. Matching the whole scan to the whole report, the way an image-captioning model matches a photo to its caption, produces a signal too blurry to pin disease to a place. DAMO's answer was organ-level fine-grained alignment — decompose the scan organ by organ, then tie each anatomical unit to the specific sentences describing it, so a line about a lesion in the right hepatic lobe trains against the liver rather than against the entire abdomen. A second component handles the fact that radiologists write the same finding a dozen different ways, learning that "mild hepatomegaly" and "enlarged liver, mild" are the same observation without anyone normalizing the text by hand.
Why It Matters
For most of the past decade, radiology AI has meant specialization. One model finds lung nodules. Another flags strokes. A third catches wrist fractures. Each needs its own painstakingly annotated training set, each is blind to anything outside its narrow slice of pathology, and each is sold separately, often on a per-scan fee. The practical result is that a hospital wanting broad coverage would need to license and run dozens of products against the same image — which is a large part of why clinical AI adoption has moved so much slower than the demo reels suggested it would.
RADAR is an argument that the narrow-model era was a constraint of technique, not a law of nature. DAMO's own history makes the point: the group produced five Nature Medicine papers on single-disease detection before this one, including PANDA, a 2023 pancreatic cancer screener trained on non-contrast scans from 3,208 patients that later drew an FDA breakthrough device designation and has since been used to screen more than 70,000 people in Chinese hospitals. RADAR is the same research line arriving at a different conclusion — that one generalist reading the whole abdomen beats a shelf full of specialists reading slivers of it.
The timing lands against a workforce problem that will not resolve itself. Researchers project a global shortfall of roughly 750,000 radiologists by 2050, and the shortage is less about a lack of senior experts than about turnaround time, geographic access, and the way expertise clusters in urban academic centers. Scans pile up in places that have scanners but not readers.
Which is why the reader-study result that matters most is not the one about beating 23 of 26 doctors. It is what happened when radiologists used RADAR alongside their own judgment instead of being measured against it. Disease-detection sensitivity rose by roughly 10 percentage points. Reading time per case fell by more than 30 percent. And junior radiologists working with the model reached accuracy comparable to senior colleagues working alone. A tool that lifts the floor while speeding up the ceiling addresses both halves of the access problem at once — which is a more interesting claim than any benchmark victory.
Reaction
The response from clinicians has been considerably cooler than the numbers might suggest, and not because anyone disputes the arithmetic. Medical AI has spent the past year running into a wall of professional skepticism, and the skepticism is specific rather than reflexive.
A survey of 355 U.S. physicians and nurses conducted in March 2026 found that daily AI use among clinicians had tripled over the previous year — and that their list of objections had grown right alongside it. Roughly 74 percent worried the technology would erode clinical skills. About the same share distrusted outputs because of hallucinations. Around 72 percent were uneasy that advertiser-driven business models might bend medical recommendations. Notably, imaging and diagnostics sit on the acceptable side of that line for most clinicians, precisely because performance can be checked against a ground truth: the scan was read correctly or it was not. RADAR lands in the one corner of medical AI where doctors are already willing to look.
The regulatory picture is blunter. RADAR has no FDA clearance, and any U.S. hospital deploying it for clinical diagnosis would be operating outside the rules. There is no obvious predicate device for a 146-disease scope, which points toward the De Novo pathway — a review that typically runs five to eight months once it starts, and it has not started. European CE marking under the EU AI Act, which treats multi-disease diagnostic systems as high-risk, would be a separate submission entirely. The weights are also released under a non-commercial license, so research use is free but clinical productization is not.
Then there is the question of who the training data represents. RADAR's 420,000 scans came from Chinese clinical centers concentrated in Zhejiang Province. Validation across eight sites addresses institutional variation within one country; it does not establish how the model performs on demographically different populations, and foundation imaging models have previously shown measurable performance gaps along racial and sex lines when moved outside their training distribution. A hospital in Boston or Berlin should treat the published accuracy as an upper bound until someone checks.
What's Next
Prospective evidence is the missing piece, and it is already in motion. A clinical trial registered as NCT07040358, sponsored by the First Affiliated Hospital of Zhejiang University with DAMO Academy, is designed to produce the real-world outcome data that retrospective validation cannot substitute for. Until those results are published, RADAR remains a rigorously tested research instrument rather than a clinical tool.
The team is explicit that the method is not tied to the abdomen. Zhang Ling, the senior algorithm expert on the project, has said the organ-level alignment approach should transfer to chest CT, brain MRI, and retinal fundus photography as clinical data partnerships expand. That is a claim about architecture rather than a shipped product — nothing outside contrast-enhanced abdominal CT has been validated — but it suggests RADAR is a template more than a destination.
The more immediate consequence may be competitive. A generalist that clears the expert bar, survives peer review at Science, and costs nothing to download is now the reference point against which every narrow commercial detector gets measured. That does not put those products out of business tomorrow; they have clearance, integration, and liability coverage that RADAR does not. But the question hospitals ask during procurement has quietly changed from which single-disease tool to buy into why they are buying single-disease tools at all.
Closing Thoughts
There is a version of this story that reads as a straightforward win — capable technology released freely, expertise democratized, everyone better off. The reporting around the release mostly reads that way. But open weights and clinical safety are different properties, and conflating them is how medical AI gets itself into trouble.
What RADAR actually demonstrates is subtler and, in a way, more encouraging. The performance did not come from a bigger model or more compute. It came from noticing that hospitals have spent decades generating millions of expert annotations as a byproduct of ordinary clinical work, and building a training method that could finally read them at the right granularity. The insight was about where the supervision already lived, not about scale. That is the kind of advance that transfers — and it is available to any research group willing to do the anatomical alignment work on their own archives.
The harder part was never the model. It is the eight months of regulatory review, the demographic validation nobody has funded yet, the PACS integration engineering, and the question of who is liable when a generalist flags 145 findings correctly and misses the one that mattered. Those problems do not have open-source solutions. But it is easier to solve them once, around a capability everyone can inspect, than separately around forty proprietary black boxes — and that may end up being the release's most durable contribution.
한글 요약
알리바바의 연구 조직 DAMO 아카데미가 9월 18일 Science에 논문을 내고, 조영증강 복부 CT 한 장에서 18개 장기에 걸친 146가지 질환을 한 번에 판독하는 AI 모델 'DAMO RADAR'를 공개했습니다. 코드는 GitHub에, 가중치는 Hugging Face에 함께 올려 누구나 내려받을 수 있게 했습니다. 학습에는 병원이 이미 보유한 복부 CT 42만 건과 여기 딸린 판독 소견서를 그대로 썼습니다. 소견서를 장기 단위로 쪼개 영상의 해당 부위와 짝지은 1,500만 쌍이 학습 데이터가 됐고, 별도의 수작업 라벨링은 없었습니다.
8개 병원에서 실제 검사 약 4만 건으로 검증한 결과 146개 소견 전체의 평균 AUC는 0.913이었습니다. 26명의 영상의학과 전문의와 맞붙은 판독 비교에서는 이 가운데 23명보다 평균 정확도가 높았습니다. 다만 더 눈여겨볼 대목은 대결이 아니라 협업 쪽입니다. 의사가 이 모델을 보조 판독자로 함께 썼을 때 질환 검출 민감도가 약 10%포인트 올랐고, 판독 시간은 30% 넘게 줄었으며, 저연차 전문의의 정확도가 단독 판독하는 고연차 수준에 근접했습니다. 2050년까지 영상의학과 전문의 약 75만 명이 부족할 것이라는 전망을 감안하면, 숫자 자체보다 이 지점이 현실적인 의미가 큽니다.
물론 곧바로 진료 현장에 쓸 수 있는 단계는 아닙니다. FDA 승인이 없고, 146개 질환을 아우르는 범위에 대응하는 기존 승인 기기가 없어 De Novo 경로를 거쳐야 할 가능성이 큽니다. 심사만 보통 5~8개월이 걸리는데 아직 시작조차 하지 않았습니다. 학습 데이터가 저장성 중심의 중국 의료기관에 몰려 있어 다른 인구집단에서의 성능도 따로 확인해야 합니다. 저장대 제1부속병원이 주관하는 전향적 임상시험(NCT07040358)이 진행 중이고, 연구진은 같은 방식을 흉부 CT·뇌 MRI·안저 촬영으로 넓힐 수 있다고 밝혔습니다. 모델을 공개했다는 사실과 임상에서 안전하다는 사실은 여전히 다른 이야기입니다.
참고
Science — DAMO RADAR 논문 · South China Morning Post · Tech Times · GitHub — damo-radar · Nature Medicine — PANDA (2023)