Decoupling Language Guidance from Backbones for Text-Guided Medical Segmentation

The paper introduces BTHA, a backbone-transferable hierarchical adapter framework for text-guided medical image segmentation. It defines a stable feature-level interface over multi-scale visual features and text representations, allowing language guidance to transfer across convolutional and transformer-based vision encoders and different language encoders. BTHA combines hierarchical coarse-to-fine supervision with a Scale-Adaptive Gated Semantic Guidance adapter that regulates textual injection by resolution and suppresses redundant cross-modal responses through channel recalibration. The abstract reports improvements over strong baselines on four public datasets with modest computational overhead, but provides no dataset names or numerical results.