Alibaba’s DAMO Academy has published two medical imaging models that read ordinary CT scans, one written up in Nature Medicine and the other in Science.
DAMO EAGLE looks for oesophageal cancer in standard non-contrast chest CT — scans taken for other reasons — and reports 90% sensitivity for cancer, 52.5% for precancerous lesions and 99.2% specificity. DAMO says it flagged malignant lesions 9 to 21 months before standard clinical pathways, validated on over 80,000 patients across 12 medical centres in three countries. DAMO RADAR reads contrast-enhanced abdominal CT and identifies more than 146 conditions from a single scan, with an area under the curve of 0.913 on nearly 40,000 real-world clinical scans. RADAR has been open-sourced on GitHub.
That matters because the input is a scan the patient was already having. Screening programmes for oesophageal cancer are thin in most countries, and a model that reads existing chest CT is a different economic proposition from one that requires its own test.
The figures are Alibaba’s own, from its own papers, and are not independently replicated. They also vary between Alibaba’s two write-ups of the same work: on how much RADAR helps a human reader, the Alizila version reports a 10% sensitivity gain, while the Alibaba Cloud version says the model matched expert radiologists. Both put the reduction in reading time at over 30%. Alibaba’s announcement is here.