Alibaba's DAMO ships two cancer screening models

Alibaba’s DAMO Academy unveiled two medical imaging models. DAMO EAGLE works on standard non-contrast chest CT with no contrast agent, screening for esophageal cancer and malignant precancerous lesions, and was published in Nature Medicine. It was validated on more than 80,000 patients across 12 medical centers in three countries (China, Czech Republic and Australia), reaching 90% sensitivity for esophageal cancer and 52.5% for precancerous lesions in opportunistic screening. In real world clinical validation on 35,402 patients, EAGLE-Plus reached 99.2% specificity, and flagged malignant lesions 9 to 21 months earlier than standard clinical pathways.

In a reader study, doctors’ sensitivity improved by up to 22.7% with specificity around 12.1%, and in clinical simulation triage roughly tripled the endoscopic detection rate while cutting healthcare costs by 37.5% to 64.5%.

DAMO RADAR is the more general one, published in Science. It is a generalist contrast-enhanced abdominal CT model identifying 146 conditions across 18 organs from a single scan, evaluated on nearly 40,000 real clinical scans at an AUC of 0.913. With RADAR assistance, radiologist sensitivity rose 10% and reading time fell more than 30%. The method uses organ-level fine-grained alignment, decomposing 3D reconstruction into anatomical units, with adaptive contrastive learning and no manual pixel-level annotation.

RADAR is open sourced on GitHub under the Alibaba DAMO Academy organisation, which is the part with the widest reach. EAGLE’s weights are not released in this announcement.

The framing is a programme rather than a single launch. DAMO already has models for pancreatic, gastric, colorectal and liver cancers, with breast and lung in development, and the stated goal is to cover the top seven leading causes of cancer death in China from a single non-contrast CT.

Two honest limits: EAGLE-Plus is described as an upgraded version of the framework rather than the released EAGLE, so treat the 99.2% specificity as a separate validation from the model being published. And the post discloses no parameter counts, architecture details, license terms or API endpoints. The wider caveat is that China accounts for roughly half of global esophageal cancer cases and deaths, and no comparison against named commercial CT CADe baselines is given.

(Original source: Alibaba Cloud)