2026-09-27
Alibaba DAMO's RADAR, trained on 400K+ CT reports with no manual labels, scores 0.883 AUC zero-shot on a US test set and lifts radiologist sensitivity ~10%, per a Science study.
Radiology AI models are usually narrow specialists: one model for one finding, retrained whenever the hospital or organ changes. Abdominal CT is particularly hard to generalize because it touches a dozen-plus organs, liver, pancreas, bowel, kidneys, each with its own normal and abnormal patterns, and manually annotating lesions at that scale is impractical. This paper from Alibaba DAMO Academy and Zhejiang University asks whether a single model can learn a broad, multi-organ diagnostic skill directly from the free-text reports radiologists already write, without any manual lesion labeling.
RADAR treats each CT scan and its radiology report as a paired image-text sample and trains on that pairing directly, skipping bounding boxes or segmentation labels entirely.
The pipeline runs on more than 400,000 contrast-enhanced abdominal CT exams and 15 million region-level image-text pairs, covering 18 anatomical structures and 146 findings.
The headline test is cross-institution transfer. A model trained entirely on Chinese hospital data was evaluated zero-shot, with no fine-tuning, on Stanford AIMI's public MERLIN test set:
| Setting | AUC |
| RADAR, zero-shot on MERLIN | 0.883 (average) |
| RADAR+, trained from scratch on Merlin-CT-Train | 0.888 (anatomy) |
| RADAR+, fine-tuned from RADAR weights | 0.918 (anatomy) / 0.876 (all) |
Per-finding results split sharply by lesion type: structurally obvious findings like abdominal aortic aneurysm (0.99), bowel obstruction (0.97), and pleural effusion (0.96) score high zero-shot, while subtle or diffuse findings like fracture (0.68) and atelectasis (0.71) lag well behind. In a reader study with 26 radiologists, RADAR assistance raised diagnostic sensitivity by roughly 10%.
Two things hold at once here: a report-supervised pipeline with no manual labeling can produce a model covering a dozen-plus organs and over a hundred findings, and that model transfers across hospitals and countries instead of overfitting to its training site. That combination is a practical path to scaling medical imaging AI past its most expensive bottleneck, expert annotation. It's also unusually open for a Science paper: weights, training code, and preprocessing scripts are all public on GitHub, Hugging Face, ModelScope, and Zenodo.
The Science article itself sits behind a paywall; this summary is built from the abstract plus the team's official code repository and documentation, not the full published text, so some methodological nuance may be missing. From what's available: training labels come from an LLM parsing raw reports with no expert re-verification, which can bake in systematic errors from either report-writing style or parsing mistakes. The [1, 1, 5] mm resampling is coarse along the z-axis, raising doubts about sensitivity to small or early lesions. The gap between high-AUC structural findings and low-AUC subtle findings (fracture, atelectasis) shows the generalist label doesn't mean uniform competence yet. The authors themselves state the model is for research use only and needs prospective clinical studies before deployment.