News · 2026-09-19

Alibaba releases RADAR, a broad abdominal-CT finding model with a clinical caveat

Alibaba's DAMO Academy has released RADAR, an anatomy-aware vision-language model for contrast-enhanced abdominal CT that the authors report achieves a mean AUC of 0.913 across 146 imaging findings. The result is notable for broad report-supervised coverage and an open research release, but it is not evidence that a model can autonomously diagnose patients or that it has regulatory clearance.

Key facts

• RADAR means Rapid Abdominal Diagnosis with AI and Radiology and is documented in the official model card.
• The authors report training on 424,911 examinations with 1.5 million image-text pairs.
• The Science/AAAS release reports 18 anatomical structures and 146 findings.
• A 26-radiologist reader study reported an approximately 10 percent sensitivity increase with assistance.

RADAR is not a chatbot looking at arbitrary scans. Its code takes volumetric NIfTI CT input, uses anatomical masks and organ-level crops, aligns regional image representations with report text and returns positive scores for named organ-finding pairs. The inference guide includes examples such as appendicitis, bowel obstruction, gallstones and aneurysm. The model is closer to a large structured checklist tied to each anatomical region than to a general physician.

Its training mechanism addresses a familiar medical-data bottleneck. Rather than requiring a clinician to draw a fresh label around every target, the system learns from examinations paired with radiology reports, then aligns the relevant region to the reported finding. The authors describe more than 15 million anatomy-specific pairs. This makes broad coverage feasible, but it also means model behavior can inherit reporting practices, institutional prevalence and the distinctions that reports do or do not state clearly.

The 0.913 result is an AUC, a ranking measure, not a percentage of patients correctly diagnosed. External-center performance was reported at 0.895 and emergency-CT performance at 0.904. The reader study is the more clinically legible number, yet it remains a controlled assistance experiment rather than proof that treatment or patient outcomes improve.

The release is open in pieces. Code is Apache-2.0 licensed, while the model card and auxiliary dataset use CC BY-NC-SA 4.0. The original 424,911-examination training corpus is not released; auxiliary masks total roughly 11 GB, but they are not the scans. Because the checkpoint files' total download size is not stated in the supplied primary materials, this report does not estimate it; the VRAM requirement is likewise unstated.

The strongest counterargument is clinical, not algorithmic: retrospective benchmarks and a reader study cannot establish performance under changing scanners, populations, prevalence, workflow pressure or liability. No public NMPA clearance was identified. The sensible so-what is that report-supervised vision transformers can now be evaluated across a much larger diagnostic inventory, but each score still needs prospective validation and a clinician responsible for the decision.

Primary source, verified: read the paper →

Ground Truth. (2026, September 19). Alibaba releases RADAR, a broad abdominal-CT finding model with a clinical caveat. Ground Truth. https://groundtruth.day/news/radar-abdominal-ct-model-release.html

@misc{groundtruth:radar-abdominal-ct-model-release,
  title  = {Alibaba releases RADAR, a broad abdominal-CT finding model with a clinical caveat},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/radar-abdominal-ct-model-release.html}
}

Topics: healthcare · medical-imaging · open-source · computer-vision · research