AGI最难一战，竟在医院！中国AI登上Science，医生不怕失业还催着上线
Ranking
Overall
74
Content
95
Popularity
26
Observed public metrics from 1 member.
Merged summary
TL;DR - Alibaba DAMO Academy’s open-source RADAR model uses vision-language learning to detect 146 conditions across 18 abdominal structures from CT scans. Reported multi-hospital results suggest general-purpose imaging AI can improve diagnostic sensitivity and reading speed while supporting, rather than replacing, radiologists.
- RADAR achieved a mean AUC of 0.913 on nearly 39,000 internal cases, 0.895 on more than 24,000 cases from eight external hospitals, and 0.904 on 27,000 emergency cases.
- Its organ-level fine-grained alignment pairs individual CT organs with corresponding report text, avoiding expensive slice-by-slice annotations and reducing irrelevant signals.
- In comparisons involving 26 radiologists, RADAR outperformed 23; AI assistance raised clinicians’ sensitivity by about 10% and reduced average reading time by over 30%.
- The model, code, and framework are open source, and the approach may transfer to other imaging modalities such as MRI, PET, and ultrasound.
Sources (1)
AGI最难一战，竟在医院！中国AI登上Science，医生不怕失业还催着上线
Public signals
OpenAlex citations 0 · Semantic Scholar citations 0 · Semantic Scholar influential citations 0
TL;DR - Alibaba DAMO Academy’s open-source RADAR model uses vision-language learning to detect 146 conditions across 18 abdominal structures from CT scans. Reported multi-hospital results suggest general-purpose imaging AI can improve diagnostic sensitivity and reading speed while supporting, rather than replacing, radiologists.
- RADAR achieved a mean AUC of 0.913 on nearly 39,000 internal cases, 0.895 on more than 24,000 cases from eight external hospitals, and 0.904 on 27,000 emergency cases.
- Its organ-level fine-grained alignment pairs individual CT organs with corresponding report text, avoiding expensive slice-by-slice annotations and reducing irrelevant signals.
- In comparisons involving 26 radiologists, RADAR outperformed 23; AI assistance raised clinicians’ sensitivity by about 10% and reduced average reading time by over 30%.
- The model, code, and framework are open source, and the approach may transfer to other imaging modalities such as MRI, PET, and ultrasound.