Alibaba Group’s research arm DAMO Academy has published a generalist medical imaging model called RADAR that its authors say can detect 146 abdominal clinical findings — including malignant tumours — from contrast‑enhanced computed tomography (CT) scans. The team released code and pre‑trained weights and published results in the journal Science, reporting strong diagnostic metrics and a radiologist comparison.
What DAMO RADAR is and what the team reports
The RADAR model is described by DAMO Academy as a vision‑language system trained on abdominal CTs and linked radiology reports. According to the research materials and the lab’s public repository, RADAR was trained on a large dataset of contrast‑enhanced abdominal CTs paired with image‑text data. Multiple outlets and the DAMO repository report the training set size as over 400,000 scans; one summary cites the figure 424,911.
Key performance figures the authors report:
- RADAR identifies 146 clinical findings spanning 18 abdominal anatomical structures.
- The model achieved a mean area under the curve (AUC) of 0.913 across 146 findings on roughly 40,000 real‑world examinations used for evaluation.
- On a subset of more than 27,000 emergency CT scans the AUC was reported as 0.904 by one summary of the paper.
- In a reader study involving 26 radiologists, RADAR’s average performance exceeded that of 23 participants; with RADAR assisting readers, missed diagnoses were reportedly reduced by 10% and reading time fell by more than 30%.
What was released and licensing caveats
The DAMO team published code and model assets. Sources that inspected the official repository state the distribution is split by license:
- Code for training and inference: Apache‑2.0 (per the repository README), which permits commercial use.
- Pre‑trained weights and auxiliary data: CC BY‑NC‑SA 4.0 (attribution, non‑commercial, share‑alike) per the public distribution on Hugging Face.
One independent commentator noted this means the procedural recipe for building RADAR is open, while the ready‑to‑use pre‑trained weights are restricted to non‑commercial use under the CC BY‑NC‑SA 4.0 terms. That distinction matters for hospitals, startups and cloud providers planning to deploy or resell inference services.
How RADAR’s approach differs from many previous models
RADAR’s authors link images to free‑text clinical reports rather than relying on manual, per‑image annotation. According to the repository description and reporting, the vision‑language training aligns CT volumes with report text and converts CT data into three‑dimensional anatomical units for more precise report alignment. The team positions this as a scalable method to cover many findings without labor‑intensive pixel‑level labelling.
Context and immediate implications
If reproduced independently, a generalist model that reads hundreds of thousands of CTs and flags diverse findings could reduce the need for dozens of disease‑specific models. The DAMO team suggests the method could extend to other imaging modalities. The published AUCs—0.913 mean across 146 findings—are substantially higher than the single‑task baseline figure one commenter cited (0.776) on the same benchmark, but that is an internal comparison and requires external validation.
What remains unresolved and what to watch for next
- Regulatory clearance and clinical integration: published model performance in a paper and repository does not equate to medical device approval in jurisdictions such as the United States; the announcement and repository materials do not document regulatory submissions or approvals.
- External reproduction: the author community flagged that independent remeasurements on data from other countries are needed to validate generalization beyond the reported test sets. One commentator suggested independent reproductions would surface within weeks of release.
- Commercial terms: while code is Apache‑2.0, the CC BY‑NC‑SA 4.0 license on weights restricts direct commercial use—an important operational constraint for companies and cloud providers seeking to offer RADAR‑based services.
Practical checklist for hospitals or developers evaluating RADAR today
- Confirm which assets you need: training code vs pre‑trained weights. Apache‑2.0 code can be used commercially; Hugging Face weights carry CC BY‑NC‑SA 4.0 restrictions.
- Run local reproduction tests: obtain the evaluation scripts from the repository and measure RADAR on in‑house, de‑identified CT datasets before any clinical trial or deployment.
- Assess regulatory pathway early: consult local medical device and health data regulators about evidence required to use RADAR for clinical decision support in your country.
- Plan workflow integration: test how RADAR outputs integrate with PACS and reporting systems and whether the model’s findings map to your radiology report taxonomy.
- Document performance drift checks: establish periodic re‑evaluation on new local cases and emergency CT subcohorts to track AUC and false negatives.
Value added beyond the press release
This article compares reported licensing—Apache‑2.0 for code versus CC BY‑NC‑SA 4.0 for weights—and highlights practical implications: open procedural reproducibility but restricted commercial reuse of the published pre‑trained model. It also gives a short operational checklist hospitals and developers can follow immediately to validate and plan regulatory and integration steps prior to any clinical use.
Sources and attribution
The technical and performance claims in this article are drawn from DAMO Academy’s public repository and the RADAR paper reported in Science and summarized by multiple outlets. Specific reported figures—”mean AUC 0.913 across 146 findings,” “about 40,000 real‑world examinations,” “trained on over 400,000/424,911 scans,” “outperformed 23 of 26 radiologists,” and the licensing split (Apache‑2.0 code; CC BY‑NC‑SA 4.0 weights)—are attributed to the project’s public materials and independent commentary cited in reporting.