RADAR+：扩展诊断范围

To extend RADAR beyond organ- specific tasks, we introduced a global image- text alignment module in RADAR+, enabling the model to associate whole- volume CT representations with complete radiology reports. This design is particularly useful for diagnostic scenarios in which the relevant pathology is not confined to a predefined anatomical structure or may involve multiple distributed regions. In the vision branch, a learnable global query token aggregates spatial features from the entire CT feature map through a cross- attention mechanism, producing a global visual representation of the scan. In the text branch, the full radiology report is encoded by the text encoder to obtain a corresponding global textual representation. …