Alibaba's Damo Academy Releases "RADAR," a Medical Vision-Language Model Capable of Identifying Approximately 150 Abdominal Diseases from CT Scans

Abdominal CT examinations present challenges, as the complexity of anatomical structures means that overlooking subtle abnormalities can lead to serious issues. RADAR was trained using over 400,000 contrast-enhanced abdominal CT scans and 15 million image-text pairs containing anatomical information. By adopting a method that learns directly from clinical reports, it reduces the burden of manual annotation (the process of labeling data) during the training stage, which has been a challenge in conventional AI development.

In performance evaluations, RADAR achieved an AUC (Area Under the Curve) of 0.913 across 146 imaging finding diagnoses within an internal cohort. Evaluations by external centers also showed AUCs ranging from 0.874 to 0.912, demonstrating high generalization performance. In validation based on pathological findings for cancer, it showed high performance with AUCs of 0.891–0.984 across four types: liver, pancreas, stomach, and large intestine. Additionally, it maintained an AUC of 0.904 for acute abdominal diseases, which were excluded during initial training.