One AI Just Read 146 Findings From a Single CT Scan

Frontier Signal · Deep Dive · AI & Investing · 6 min read · Monday, 28 September 2026

A radiologist reading an abdominal CT scan hunts through eighteen organs and dozens of possible diseases, but most AI tools built to help are trained to find exactly one of them. A liver-cancer model does not look for appendicitis, and a stroke-triage tool does not look for a bowel obstruction. Building a new model per disease is why hospital AI has stayed narrow even as the underlying technology improved.

Alibaba's DAMO Academy tried a different approach: train one model on the reports radiologists already write, rather than on diseases hand-labelled one at a time, and let it learn all 146 findings at once.

The result, published in Science in September, is a model called RADAR. In the reader study, the paper reports expert-level performance and shows that RADAR assistance increased the diagnostic sensitivity of 26 radiologists by about 10 percent. Alibaba then released the code and model weights publicly for research use.

0.913 AUC ACROSS 146 FINDINGS · NEARLY 40,000 REAL-WORLD SCANS · RELEASED FREE NINE DAYS LATER
Figures are from the peer-reviewed Science paper and Alibaba's public release; the retrospective benchmark and the reader study are separate evaluations, not one sample.

DAMO Academy, Alibaba's research arm, built RADAR with clinicians at Zhejiang University's First Affiliated Hospital, publishing the work in Science on 18 September 2026. The team trained it on 420,000 CT exams and their existing reports, then expanded that into 15 million organ-level image-text pairs by matching each report to the organ it described.

Tested on nearly 40,000 real-world exams, RADAR reached a mean AUC of 0.913 across 146 findings spanning 18 organs. On external validation across eight outside centers and roughly 24,000 more cases, accuracy held in the high 0.8s to low 0.9s; on emergency-department cases it had never been tuned for, the figure was 0.904.

The team also ran a reader study with 26 radiologists from 14 hospitals. Assisted reading raised diagnostic sensitivity by about 10 percent and cut reading time by roughly 31 percent.

Most medical-imaging AI starts with a fixed list of things to find, then needs a radiologist to draw boxes around thousands of examples of each one before the model can learn it. That does not scale past a handful of diseases per model, which is why hospital AI today is a shelf of narrow tools rather than one broad one.

RADAR skips the manual labelling. It treats the report that already comes with every scan as a ready-made label, then aligns regions of the CT volume with that report's language. Because an abnormal finding often occupies a tiny fraction of a full scan, the team first split each volume into its component organs, turning one whole-scan needle-in-a-haystack problem into eighteen smaller, tractable ones.

In plain English: instead of paying people to point at diseases one at a time, RADAR learned by reading the paperwork a hospital already produces, organ by organ, at a scale no manually labelled dataset could reach. It runs on established open tools for vision-language learning and image segmentation, not Alibaba's own Qwen models.

Alibaba's own history makes the contrast easy to see. DAMO has already built and deployed narrow single-disease tools: PANDA for pancreatic cancer, which holds FDA Breakthrough Device status and has screened more than 70,000 patients in China; LiON for liver cancer, which caught missed metastases in a real hospital trial two months before RADAR was published; and iAorta for aortic emergencies. Each needed its own dataset and its own multi-year validation.

RADAR is DAMO's bet that the one-disease-one-model pattern does not have to continue. If a single model trained on ordinary hospital paperwork can cover 146 findings at research-grade accuracy, the bottleneck in medical AI shifts from labelling data toward something closer to a software rollout, a faster kind of problem.

The same one-model-instead-of-many wager has already paid off elsewhere: weeding robots that used to need a new brain for every crop now run on a single foundation model too.

DAMO is not shy about deployment: its narrower tools run in Chinese hospitals, and the company has piloted multi-disease CT screening in Singapore, Japan and Saudi Arabia. A research-only license puts a 146-finding model into the hands of hospitals and labs years before a cleared commercial version could exist, useful for independent testing, but it blurs the line between research evaluation and clinical use.

Alibaba (NYSE: BABA; HKEX: 9988) is the parent of DAMO Academy, and this is not an earnings story: no earnings call mentions DAMO, RADAR or healthcare AI by name, despite accelerating cloud growth and a modest AI-linked revenue run-rate. This looks like a research and reputation programme apart from the Cloud and Qwen businesses investors actually price.

Aidoc, privately held, won FDA clearance in January 2026 for a tool screening one abdominal CT for fourteen critical findings such as bowel obstruction and organ injury, one of the first clearances anywhere for a multi-condition AI reader. Fourteen cleared findings against RADAR's 146 uncleared ones is the real trade-off: breadth versus regulatory maturity and hospital contracts.

GE HealthCare (NASDAQ: GEHC), Siemens Healthineers and Philips build the scanners and workflow software radiologists actually use, and all three already sell cleared AI-enabled imaging tools. None has commented on RADAR. A free research model does not threaten their installed base today, but it raises the bar for what a narrow paid tool must justify.

Beyond Aidoc, Viz.ai focuses on stroke detection and alerting the right specialist fast, more a workflow product than a diagnostic-accuracy race. Annalise.ai and Qure.ai work narrower still, on head CT and chest X-ray, often in health systems that cannot support RADAR's compute needs. None offers RADAR's scope; all offer something it lacks, a regulatory clearance and a paying hospital.

RADAR is peer-reviewed in Science, more than most AI-health claims can say. But the headline metrics come from different evaluation layers and should not be collapsed into one result.

The 0.913 AUC comes from the retrospective evaluation; the reader study is a separate experiment. The safe read is that the model performed strongly at scale and also improved human readers in that study, not that every comparison applies across the whole benchmark.

No outside lab has yet published an independent reproduction, and training and most validation came from Chinese hospitals. The release is non-commercial too: the CC BY-NC-SA 4.0 license means a vendor cannot turn these weights into a paid product without a separate arrangement.

The same caution applies to any confident AI output before outside scrutiny catches up, the subject of our look at why AI hallucinates and cannot fully stop.

Would this model's accuracy hold on our patients, in our hospital, or only on the patients it was trained and tested on?

That is the question a strong AUC on Chinese hospital data cannot answer by itself, and it is the one a health system should demand evidence for before treating RADAR's numbers as its own.

It read abdominal CT scans for 146 distinct findings across 18 organs at once, reaching a mean accuracy score of 0.913 on nearly 40,000 real-world hospital scans, published in Science on 18 September 2026.

In the reader study, RADAR showed expert-level performance, and radiologists using it as an assistant improved sensitivity by about 10 percent. That study was separate from the retrospective benchmark, so the two results should not be merged.

Anyone can download the code and weights for free, but the license is research-only, and DAMO's own documentation says further studies are needed before clinical use.

No. RADAR is built on separate open-source tools for vision-language learning and medical image segmentation.

Almost all training and testing data came from Chinese hospitals, with only one external Western dataset checked, so it is not yet established how the model performs on more diverse patient populations.

Frontier Signal explains frontier technology in plain English. Company-reported and vendor-reported claims should be independently verified. This is general information, not investment advice.