RADAR：用于腹部CT诊断的视觉-语言基础模型

into textual features. The model was trained on ~15 million anatomy-wise image-text pairs. These pairs were derived from the Rapid Abdominal Diagnosis–CT (RAD-CT) cohort by anatomically decomposing volume-wise image-text data from more than 400,000 cases at a single high-volume center (Fig. 1, B and C). Once trained, the joint feature space enables prompting-based diagnosis without any fine-tuning. For every diagnostic query, two textual prompts are designed: a positive prompt describing its presence (for example, “colitis”) and a negative prompt describing its absence (for example, “no evidence of colitis”) (Fig. 1D). Both prompts are encoded by the text encoder and compared with the corresponding anatomy-specific image feature using cosine similarity. …