Alibaba’s research unit, Damo Academy, has released to the public a powerful artificial intelligence (AI) model that reads computed tomography scans of the abdomen. The model is called Damo Radar. The institute said it was built to study contrast-enhanced scans. It covers 18 organs in the belly and is meant to pick out nearly 150 conditions, including cancers and other abnormalities such as malignant tumours, meaning growths that can invade nearby tissue.
The system is a vision-language model. This means it learns from pictures together with written language, in this case the clinical reports that doctors already write beside the scans, rather than from extra hand-drawn labels on every image. In almost 40,000 real examinations it reached an average AUC of 0.913 across 146 clinical findings. AUC is a score of how well a test separates people who have a condition from people who do not. A score of 1.0 would be perfect separation; 0.5 would be no better than a coin toss.
How doctors fared beside the model
The researchers called Damo Radar the world’s first expert-level generalist medical imaging model and said the same training method might later be used on other kinds of scans. In a comparison with 26 radiologists, doctors who specialise in reading medical images, drawn from several hospitals, the model’s average accuracy was higher than that of 23 of those doctors. When the doctors used the model as an aid, they improved their ability to avoid missed diagnoses by 10 percent and cut the time required by more than 30 percent, Damo said.
The study is published in Science and the code is also published. A hospital affiliated with Zhejiang University took part with Alibaba. The release follows earlier Damo tools meant to help find pancreatic, stomach and colorectal cancers and tears in the aorta, the body’s main artery. In April the unit released Coca, which it said was more sensitive than radiologists at spotting early colorectal cancer on CT, in work with Guangdong General Hospital and other Chinese institutions.
English summary of the Alibaba news release in Chinese: the model is aimed at contrast-enhanced abdominal CT diagnosis, can identify more than 146 conditions in a single pass, has for the first time reached the accuracy of specialist radiologists, and is now fully open-sourced.