Alibaba AI reads one CT scan for many findings

On September 18, 2026, Alibaba DAMO Academy publicly announced that its vision-language model RADAR (Rapid Abdominal Diagnosis with AI and Radiology) had been published in the journal Science and released it as open weights (FACT — SCMP "the institute said on Friday", IT之家 16:00 Beijing relay of the DAMO announcement, Gate News flash). The peer-reviewed paper, "An expert-level generalist AI for abdominal CT diagnosis" (Qi Zhang, Jianpeng Zhang, et al.; Science vol 393, issue 6817, eaec6129; DOI 10.1126/science.aec6129; print date 17 September 2026), describes a generalist AI that reads a contrast-enhanced abdominal CT scan and diagnoses 146 imaging findings across 18 anatomical structures in one go (FACT — paper abstract via PubMed/EurekAlert/Ovid).

Key reported results (FACT — peer-reviewed paper's abstract-level claims, consistently relayed by AAAS and every independent outlet):

• Mean AUC 0.913 across the 146 assessed findings on real-world hospital examinations (~39,160 consecutive exams internal test set), vs 0.776 for the best competing vision-language model — a large gap over specialist/VL baselines.
• Emergency generalization: AUC 0.904 across 27,267 emergency CT cases, despite the model not being specifically trained on emergency data (out-of-distribution acute-abdomen performance; Pandaily: "AUC remained about 0.904").
• External validation: 24,239 examinations from 8 independent hospitals, AUC 0.874–0.912 (MedicalXpress).
• Reader study: 26 radiologists from multiple hospitals (14 centers per AI/TLDR) — RADAR's average accuracy exceeded 23 of the 26; with RADAR assistance radiologists' diagnostic sensitivity rose ~10% and reading time fell ~30% (MedicalXpress: "roughly 31%"), and junior radiologists with assistance reached senior-level performance.
• Training scale: 424,911 contrast-enhanced abdominal CT examinations; 1.5 million image–text pairs; more than 15 million anatomy-specific pairs; the model learns directly from clinical radiology reports without manual annotation (FACT — EurekAlert/AAAS abstract).

Release (FACT, verified directly): code and weights published to GitHub (alibaba-damo-academy/damo-radar), Hugging Face (radar-generalist/RADAR), ModelScope (modelscope.cn/collections/DAMO_Academy/RADAR), and Zenodo (records/21271172) — all under CC BY-NC-SA 4.0, with RADAR+ checkpoints (a variant trained/fine-tuned on the external MERLIN dataset) included.

Institutional context: the work is a collaboration between DAMO Academy, the First Affiliated Hospital of Zhejiang University School of Medicine (corresponding authors Tingbo Liang, Qi Zhang, Wenbo Xiao), Hupan Laboratory (Jianpeng Zhang), and DAMO algorithm experts (Ling Zhang), with 30+ authors across Chinese medical centers. DAMO's medical-AI track record (per IT之家/Sohu relays of the announcement): three years of medical AI work, 5 Nature Medicine papers, and prior systems for pancreatic/gastric/colorectal cancer screening and aortic-dissection alerting.

• A peer-reviewed, expert-level generalist medical-imaging model entered the open-weight ecosystem. Radiology AI to date has been "one disease, one model" (lung nodule detectors, stroke flags, mammography readers). RADAR is the first model whose Science paper demonstrates expert-level, multi-organ, multi-disease diagnosis from a single abdominal CT exam — the "generalist moment" for medical imaging AI (FACT on the paper's claims; this framing is the team's and is echoed across independent coverage).
• The training paradigm changed from annotated supervision to report-derived self-supervision at scale. Instead of physicians annotating lesions slice-by-slice, RADAR learns by aligning CT anatomy with the sentences already written in hospital radiology reports — removing the manual-annotation bottleneck that made multi-disease models expensive and slow to build (FACT — EurekAlert abstract; INTERPRETATION on significance).
• Open weights with a non-commercial license entered medical AI's mainstream conversation. The release is open (weights downloadable, code reproducible) but CC BY-NC-SA 4.0 — deliberately research-oriented, with the repo stating the model is "intended for research purposes only" pending prospective clinical studies (FACT, verified license).
• A Chinese lab+hospital consortium now holds the high ground on generalist medical imaging AI in a top journal — a category where most prior landmark results involved single-disease/detector pipelines from a variety of international groups (INTERPRETATION grounded in the paper's own "generalist" positioning and DAMO's press framing).
• Human-AI reader studies moved from "does AI beat the doctor" to "does AI-assisted reading move the whole distribution" — the ~10% sensitivity gain and ~30% reading-time cut with junior radiologists reaching senior levels is the most clinically actionable finding (FACT on reported study; the staffing/product implications are INTERPRETATION).

Before (through Sep 16, 2026):

• Abdominal CT — one of radiology's hardest workloads (dozens of overlapping organs, hundreds of possible findings per study) — was served by narrow specialist AIs: one model per organ or per disease, each requiring labor-intensive annotated training data and typically validated on small, single-center cohorts.
• Competing vision-language medical models (the paper reports the best at 0.776 mean AUC across findings) trailed expert readers; no generalist system had demonstrated expert-level multi-organ performance in a peer-reviewed venue.
• DAMO Academy's medical AI output was a series of disease-specific systems (pancreatic/gastric/colorectal screening, aortic-dissection alerting — 5 Nature Medicine papers in three years per the announcement relays) with no unified generalist model.
• Open-weight medical imaging releases existed (nnU-Net/MONAI-era tooling) but as components, not as an expert-level end-to-end diagnostic model.

Change (Sep 17–22, 2026):

• Sep 17: Science prints "An expert-level generalist AI for abdominal CT diagnosis" (vol 393, issue 6817, eaec6129); AAAS issues the embargoed press release (EurekAlert, 17-Sep-2026).
• Sep 18: DAMO Academy announces the work and open-sources RADAR (GitHub + Hugging Face + ModelScope + Zenodo, CC BY-NC-SA 4.0); SCMP, Yicai, Gate News, IT之家, and others cover it within hours.
• Sep 19–22: English-language and international coverage spreads (Pandaily, AI/TLDR, aitrove, MedicalXpress, Artificial Science, NDTV Profit, developmentstoday, voi.id); the MEDLIN-indexed PubMed record appears (PMID 42752131); community Spaces appear on Hugging Face (junma/RADAR-demo, loupk/RADAR-demo, scomb2/Radar).

After:

• Any researcher can download RADAR's weights and code, run the provided inference/RAD-CT demo or evaluate on the MERLIN test set, and reproduce a 146-finding/18-organ abdominal CT read at 0.913-class AUC — under a non-commercial license.
• Radiology-adjacent builders now have a peer-reviewed reference for "report-supervised generalist imaging AI" and a concrete transfer thesis (MRI/PET/ultrasound next).
• The competitive frame for medical imaging AI shifted: the next landmark papers in this space will be measured against RADAR's numbers, and regulators now have a concrete, citable example of a generalist model that is not yet cleared for clinical use anywhere (INTERPRETATION + PREDICTION).

Vision-language learning from clinical reports (FACT — paper abstract/EurekAlert): instead of supervised lesion annotation, RADAR learns the association between imaging content and the text of radiology reports at scale. Training corpus: 424,911 contrast-enhanced abdominal CT examinations → 1.5M image–text pairs → >15M anatomy-specific pairs.

Organ-level fine-grained alignment (FACT as described by the team/paper; the mechanism is documented in the peer-reviewed work and relayed by IT之家 in the DAMO announcement): because CT signal is sparse, plain vision-language learning underperforms. RADAR converts the CT volume into 3D, decomposes it into individual anatomical structures, and aligns image content with report text at the organ level, using a large-scale contrastive-learning objective that dynamically adapts (adaptive contrastive modeling). This is the team's stated "first international use" of organ-level fine-grained alignment (COMPANY CLAIM on the "first" — the mechanism itself is in the paper).

Model components (FACT, from the open repository): the released stack includes a vision branch with a pretrained UNet checkpoint sharing the nnU-Net lineage (checkpoint_unet.pth), language encoders (BERT-base-Chinese and BERT-base-uncased checkpoints), the pretrained RADAR checkpoint (checkpoint_radar_pretrain.pth, trained on RAD-CT), and RADAR+ variants (trained from scratch on MERLIN-CT-Train, and pretrained-on-RAD-CT-then-fine-tuned-on-MERLIN). The codebase builds on LAVIS (BSD-3), nnU-Net (Apache-2.0), MONAI (Apache-2.0), and 3D-ResNets-PyTorch (MIT) — with third-party licenses retained in THIRD_PARTY_LICENSES.md (FACT, verified on repo).

Inference/evaluation workflow (FACT, repo docs): the repo ships RADAR_inference (an inference demo with the pretrained checkpoint on RAD-CT; inference and evaluation on the external MERLIN test set), RADAR_train (train RADAR/RADAR+ from scratch or fine-tune on MERLIN data), and preprocessing code for image/mask and report preprocessing of MERLIN or custom data. Checkpoints and auxiliary data (processed masks) are downloaded via download_scripts/ from Hugging Face.

What it does NOT do (FACT — repo disclaimer + independent analysis): it has no regulatory clearance as a medical device; the model card states "Further improvements and prospective clinical studies are still required before it can be used directly for clinical deployment"; it is trained/validated on Chinese patient populations and contrast-enhanced abdominal CT only; it is a research/second-reader aid, not an autonomous diagnostic product.

1. It is the flagship concrete health-AI result of this news window, anchored by peer review. Within the same week as model releases and policy actions, RADAR is the window's rare "peer-reviewed, open-sourced, expert-level" medical-claim — the numbers (AUC 0.913; +10% sensitivity; −30% reading time; external hospitals) are reproducible claims from a Science paper, not a launch marketing sheet. (FACT on the paper; significance is INTERPRETATION.)
2. It reframes medical AI from "hook the next organ" to "one model for the abdomen, then the body." The generalist framing — 146 findings across 18 structures in one read — attacks the integration bottleneck that stalls clinical adoption (hospitals can't run forty narrow models per scan), which is precisely why aitrove/Artificial Science characterize it as medical AI's "generalist moment." (INTERPRETATION grounded in independent coverage.)
3. It demonstrates the economics of report-derived supervision at scale. No manual annotation → the marginal cost of expanding to new organ systems/diseases collapses, and hospitals' existing report archives become training fuel. The transfer thesis (MRI, PET, ultrasound) is the strategic payload. (FACT on the mechanism; implications are INTERPRETATION.)
4. It opens a research-grade expert reader to the entire radiology-AI community — under a non-commercial license — including RADAR+, the MERLIN external-set evaluation harness, and three Hugging Face demo Spaces within days. For academic and non-commercial hospital work this is a free, reproducible baseline. (FACT on release artifacts.)
5. It puts a Chinese lab+hospital consortium at the center of the generalist-imaging conversation, with an internationally legible publication (Science) and fully public artifacts — a shape of China AI leadership in healthcare that differs from the model-release stories (Qwen/MiMo/StepFun) in the same window. (INTERPRETATION.)

• A single abdominal CT exam read for 146 findings across 18 structures in one pass — liver, pancreas, gastric, colorectal, and other cancers plus masses, fractures, and anomalies — at reported expert-level AUC (0.913), as a research/second-reader tool (FACT on claims; "possible" subject to the research-only caveat).
• Report-supervised training at hospital scale: institutions with CT+report archives can now follow RADAR's documented recipe (preprocess → align at organ level → contrastive pretrain → fine-tune) to build their own generalist readers without manual annotation (FACT — code and docs are public).
• Reproducible external evaluation: the MERLIN test-set pipeline and RADAR+ baselines let any research group benchmark its own model directly against the published work (FACT — repo).
• Non-commercial clinical-innovation pilots inside hospitals: research departments, radiology residencies, and startups doing non-commercial validation can deploy the model behind internal tools to test second-reader workflows (subject to CC BY-NC-SA 4.0 and institutional rules).
• A template for the "generalist imaging foundation model" research program — the team explicitly frames MRI/PET/ultrasound as natural transfer targets (COMPANY CLAIM / research thesis per zglg relay of the team's remarks).
• Not possible: commercial deployment without a separate license; any regulatory-cleared diagnostic use anywhere; use on non-contrast CT, X-ray, MR, or other body regions (currently); autonomous (radiologist-free) reporting per the model's own positioning.

1. Will the ~10% sensitivity / ~30% time-savings assist results replicate in prospective, multi-country trials — and in non-Chinese patient populations?
2. How does RADAR perform on non-contrast CT (acute settings often skip contrast) and on other vendors' reconstruction protocols beyond the 8-hospital external set?
3. Which commercialization vehicle follows — a DAMO-cleared derivative (e.g., NMPA pathway) and/or a commercial license for the current weights — and when? (PREDICTION: within 12 months a cleared derivative is likely.)
4. Do the "scaling with data, no saturation" curves hold for 1M+ exams, and does organ-level alignment transfer cleanly to MRI/PET/ultrasound as the team claims?
5. Is the "first expert-level generalist" claim contested by other groups' concurrent work (e.g., generalist whole-body or multi-modality models in submission)?
6. What do the MERLIN-set RADAR+ results look like vs the published RAD-CT-pretrained model — does fine-tuning on external data close the 0.874–0.912 band's lower end?
7. Which community effects materialize: how many research groups run the open release, and do demo Spaces/forks translate into independent validation papers?

• Days (to end of September): community activity on the release — GitHub stars/forks (63 stars/12 forks at research time — the repo went public Sep 18), Hugging Face demo Spaces, first independent replication attempts begin; press coverage consolidates around the "beats most radiologists" headline with increasing caveats about the non-commercial license and lack of clearance.
• Weeks (October–November): expect (a) the first community reproductions of AUC 0.913-class results on RAD-CT and MERLIN; (b) commentary from rival medical-imaging-AI groups contesting or benchmarking the "first expert-level generalist" claim; (c) DAMO follow-up releases — expanded findings coverage, new modalities, or a Chinese clinical-validation announcement; (d) the licensing discussion surfacing in radiology-AI procurement blogs (who can actually use this commercially?). (PREDICTION.)
• Structural watch (quarter): whether generalist report-supervised imaging models become the default research paradigm (RADAR as the reference open implementation), whether regulators begin addressing "generalist models" as a distinct device category, and whether DAMO productizes a cleared derivative that turns the research release into a commercial franchise. (INTERPRETATION + PREDICTION.)

RADAR is the week's flagship peer-reviewed health-AI story: a Science-published, open-weights vision-language model that reads a contrast-enhanced abdominal CT and reports 146 findings across 18 anatomical structures at expert-level mean AUC 0.913 — with genuine robustness evidence (0.904 on 27,267 emergency cases the model wasn't trained on; 0.874–0.912 across 8 external hospitals, 24,239 exams) and a clinically meaningful assist result (radiologist sensitivity +~10%, reading time −~30%, juniors reaching senior levels). The verified core is real and unusually clean: paper identity, training recipe (424,911 exams, 15M anatomy-specific pairs, report-derived supervision without manual annotation), release artifacts on GitHub/Hugging Face/ModelScope/Zenodo, and a consistent set of numbers relayed by AAAS. Three honest corrections every reader needs: (1) the license is CC BY-NC-SA 4.0 on both code and weights — open weights, research-only, non-commercial — so "fully open-sourced" headlines overstate; (2) this is abdominal contrast-enhanced CT, not whole-body screening, and the "20+ hospitals" figure in circulation is not directly documented (8 independent hospitals + 26 radiologists/14 centers are); (3) the model has no regulatory clearance anywhere and is explicitly research-only pending prospective clinical studies — the best use of this release is as the reproducible, peer-reviewed generalist baseline that moves medical imaging AI from "one disease, one model" to "one expert-level model per exam class," and the story the field will now argue about is who beats its numbers next.