Alibaba DAMO Academy Open-Sources DAMO RADAR Universal Medical Imaging AI
Alibaba DAMO Academy and its research partners published a study in Science introducing DAMO RADAR, an open-source universal medical imaging AI model capable of diagnosing over 146 abdominal conditions from CT scans with expert-level accuracy. The model was trained using a vision-language learning approach that eliminates the need for manual data annotations. DAMO RADAR represents a breakthrough as the first universal medical imaging model to achieve expert performance across dozens of organs and conditions simultaneously. In clinical trials, AI-assisted diagnoses improved radiologist sensitivity by 10% and reduced reading time by over 30%, helping elevate junior doctors to senior diagnostic standards. The model uses an 'organ-level fine-grained alignment' strategy to decompose 3D CT scans into anatomical units and align them with clinical text reports automatically. Across nearly 40,000 real-world imaging checks, DAMO RADAR achieved an Area Under the Curve (AUC) of 0.913 and outperformed 23 out of 26 expert radiologists in comparative trials.
## BACKGROUND
Medical imaging AI has traditionally relied on highly specialized models trained on manually annotated datasets to detect specific, single diseases. Vision-Language Models (VLMs) combine visual imagery with medical text to enable broader learning, but applying them to 3D CT scans has been challenging due to signal sparsity in volumetric imaging.