One CT Scan, 146 Diseases: A Milestone in AI Radiology
On September 18, 2026, Alibaba's DAMO Academy published a peer-reviewed study in Science describing DAMO RADAR, a generalist medical imaging AI system capable of screening 146 distinct diseases across 18 organs from a single contrast-enhanced abdominal CT scan. In a head-to-head reader study, RADAR's mean diagnostic accuracy exceeded that of 23 of 26 specialist radiologists.
Critically, the model's code is freely available on GitHub and HuggingFace under an Apache 2.0 license, with model weights released under CC BY-NC-SA 4.0 for research use. Any hospital or research institution worldwide can download and deploy it immediately.
Why This Matters
For the past decade, radiology AI has meant specialization: one model for liver lesions, another for pancreatic masses, a third for appendicitis. Each requires painstakingly annotated training data and remains blind to anything outside its narrow training distribution. The commercial result has been a fragmented market of point-solution products, each charging separately per disease class.
RADAR breaks from that architecture entirely. It covers 18 organs and 146 diseases in a single pass, validated externally at eight independent clinical centers on approximately 40,000 real-world examinations, achieving a mean AUC of 0.913. AUC measures a diagnostic test's ability to distinguish diseased from non-diseased patients: 1.0 is perfect, 0.5 is chance, and above 0.9 is conventionally considered "outstanding." RADAR sits firmly in that top tier.
The Technical Breakthrough: Organ-Level Fine-Grained Alignment
CT volumes are three-dimensional and sparse — in a typical contrast-enhanced abdominal CT, the overwhelming majority of voxels contain normal tissue, and pathology is localized and often small. Standard vision-language models (VLMs) align the whole image to the whole report, producing a training signal too coarse to localize disease reliably.
The DAMO team's solution is organ-level fine-grained alignment:
- Each CT volume is decomposed organ by organ — liver, kidneys, spleen, bowel, pancreas, and 12 other structures;
- Each anatomical unit is aligned specifically with the corresponding sentences in the paired radiology report;
- An adaptive contrastive modeling strategy handles variation in how different radiologists phrase the same finding (e.g., "mild hepatomegaly," "enlarged liver," "hepatomegaly, mild");
The result: a model trained on 420,000 contrast-enhanced abdominal CT examinations generating 15 million anatomy-aware image-text pairs — a scale historically reserved for natural-language foundation models, now applied to volumetric medical imaging. Crucially, the entire pipeline required no manual annotation beyond existing radiology reports.
Beating Most Radiologists — and Making the Rest Better
The headline comparison in the Science paper was a formal reader study against 26 specialist radiologists spanning junior residents to attending-level physicians from multiple institutions. RADAR's mean diagnostic accuracy exceeded that of 23 of the 26 physicians.
More clinically significant, however, is what happened when RADAR was used as an AI second reader alongside radiologists:
- Disease-detection sensitivity improved by approximately 10 percentage points;
- Reading time per case fell by more than 30 percent;
- Junior radiologists with RADAR assistance reached diagnostic accuracy comparable to senior colleagues working unaided.
This matters for the radiologist shortage. Researchers project a shortage of 750,000 radiologists by 2050. A tool that brings junior-physician performance up to senior levels while senior radiologists work faster addresses both capacity and speed simultaneously.
What the Model Will Not Do (Yet)
RADAR is trained exclusively on contrast-enhanced abdominal CT — a specific protocol requiring intravenous contrast before scanning. The following are outside the validated scope:
- Non-contrast CT (widely used in emergency screening and population health programs);
- Different contrast phases (arterial, portal venous, delayed);
- Pediatric imaging;
- All other body regions (chest X-ray, brain MRI, etc.).
A hospital cannot use RADAR to screen a plain-film chest X-ray, a brain MRI, or a non-contrast abdominal study. Additionally, the model weights are released under CC BY-NC-SA 4.0, which prohibits commercial deployment without a separate agreement.
What It Means for Patients
In the near term, RADAR's open-source release will accelerate testing and deployment of generalist imaging AI at hospitals and research institutions worldwide. For patients undergoing abdominal CT, the future may bring faster report turnaround, more consistent diagnostic quality (especially at community hospitals), and earlier detection of potential pathology.
AI will not replace doctors. RADAR is positioned as a "second reader" — helping physicians see faster and more accurately, not replacing clinical judgment.
Tracking your own health metrics is the first step in disease prevention. The Xiaoshu Health App can sync with Apple Health to track relevant indicators, helping you record weight, blood pressure, blood sugar, and more — building a complete personal health profile alongside regular checkups and imaging exams.