DAMO's open, free CT AI model matches top radiologists across 146 diseases

Two licences do two different jobs here. The code went up on GitHub and HuggingFace under Apache 2.0 ^[3], and the weights, which are the expensive output of any training run, went out separately under CC BY-NC-SA 4.0 ^[4]. Research use is allowed; commercial deployment needs a separate agreement ^[4]. Alibaba did not disclose what that agreement costs. The download is free for a university group ^[8], and for a hospital that bills for reads the separate agreement is the price.

So the pressure on incumbents starts as a number they have to answer. External validation ran at eight independent clinical centres on about 40,000 real-world examinations and produced a mean AUC of 0.913 ^[6]. The conventional threshold for "outstanding" clinical performance is 0.9 ^[7][18]. TechTimes describes the market RADAR lands in as a fragmented set of point-solution products, each charging separately for each disease class it covers, on per-scan fees for narrow detectors ^[9]. A procurement officer negotiating a renewal does not need the weights to cite the Science result ^[1].

The training run used 420,000 contrast-enhanced abdominal CT examinations and produced 15 million anatomy-aware image-text pairs ^[11], about 36 pairs per examination ^[15], or roughly two per evaluated organ ^[16]. Those pairs came out of radiology reports that had already been written, matched organ by organ to the region of the scan a given sentence describes; standard vision-language models align the whole image to the whole report, which is too coarse a signal to localise disease in a volume where most voxels are normal tissue ^[12]. The validation set is under a tenth the size of the training corpus ^[17].

DAMO's research team called RADAR "the world's first expert-level generalist medical imaging model" and said the training method could eventually be extended to other types of medical imaging, according to the South China Morning Post ^[13].

The commercial licence could be priced like a per-study toll, in which case the cost of reading a scan does not move and only the counterparty changes. It could also be that the research licence is the only one anybody ever uses, if hospitals and national health systems run RADAR under research cover and narrow detectors lose their premium without anyone signing anything. And 0.913 is a mean across 146 findings ^[6]. The model outperformed 23 of the 26 radiologists in the head-to-head study ^[19]. If the findings that carry the most clinical liability sit in the tail below that mean, the per-disease vendors keep selling into exactly the cases they were built for.

I'd expect the second, because the demand argument is large and the cost of trying is zero: TechTimes cites a projected shortage of 750,000 radiologists by 2050 ^[10]. DAMO put its training budget into one generalist covering 18 organs and 146 diseases in a single pass ^[5] and left commercial deployment of the weights to a separate agreement ^[4]. I'd be wrong if that licence is priced at incumbent per-scan rates, or if a per-finding breakdown puts the cancer findings below the radiologists who read against them.