AINEWS
Alibaba's research arm released the model on GitHub this week alongside a peer reviewed Science paper, and both the 0.913 score and the radiologist comparison came from the same team that built it.
Alibaba's research arm this week open-sourced a medical-imaging AI that flags 146 conditions in abdominal CT scans. In a single peer-reviewed study of nearly 40,000 real-world scans, the model hit an average AUC of 0.913, a 0-to-1 ranking score where 1.0 is a perfect classifier. The team that built the model also wrote the test, and reported that the model matched or beat most radiologists on the comparison.
The 146 findings all sit in the abdomen. The model reads contrast-enhanced CT scans, the kind where a patient is injected with a dye that brightens blood vessels and organs to make tumors and inflammation easier to spot. It is a vision-language model, a class of AI that trains on images paired with text descriptions so it can later answer questions about new images. Damo Academy trained Damo Radar on more than 400,000 contrast-enhanced abdominal CT examinations paired with 15 million anatomy-aware text descriptions, per a Vuink summary cited in SCMP's coverage.
The same week as the GitHub release and the Zenodo archive, the paper, "An expert-level generalist AI for abdominal CT diagnosis," appeared in Science (DOI 10.1126/science.aec6129). Outside labs can now download the weights and audit the result on their own CT set.
AUC 0.913 leaves a 0.087 gap to a perfect score, and the "near-perfect" framing in some same-week coverage overstated the result. The paper's average AUC is computed across all 146 findings, including common ones the model handles well and rare ones it does not, so a 0.913 average can sit on top of a handful of conditions where the model performs much worse.
SCMP's "outperformed most radiologists" line traces back to the Damo Radar research team, not an independent benchmark. The paper's lead author is Qi Zhang, with senior authors Yingda Xia and Tingbo Liang, all from Damo Academy. The radiologist panel was assembled by the same group. The people who chose the test set, the radiologist labels, and the scoring rules also built the model. A same-team finding is not the same kind of evidence as an outside head-to-head, and the comparison can still be a fair one.
The 146 conditions are all abdominal, so "generalist" here means one body region, not all of medical imaging. The model has not been cleared for clinical use by any regulator in the public receipts. Hospital deployment would still require integration with PACS, the picture archiving systems radiologists already use, a workflow for handling false positives, and a liability path for the cases the model misses.
The open-source release lets outside teams run the test. A hospital, an academic lab, or a competing team can pull the weights from the Zenodo archive and run the model against a retrospective CT set the Damo team never saw. The GitHub repository ships evaluation scripts. Community discussion on Hacker News has already surfaced the same-team question; the missing piece is a radiologist-led benchmark on hospital data the Damo team never saw.