RADAR：腹部CT的视觉-语言基础模型

Table 1. Cross-population evaluation on the public Merlin-CT-Test benchmark. The upper section compares the AUC performance of RADAR with other vision-language models, including BiomedCLIP, MedGemma, and Lingshu, on the unseen Merlin-CT data, highlighting their zero-shot generalization capabilities. The lower section presents AUC results for methods trained directly with Merlin-CT-Train data, specifically Merlin, RADAR trained from scratch, and RADAR+, where “RADAR+” denotes a fine-tuned version of RADAR on the Merlin-CT dataset. …