BTHA decouples language guidance from backbones for medical segmentation

Researchers present BTHA, a backbone-transferable hierarchical adapter framework designed to decouple language guidance from specific vision and text encoders in medical image segmentation.

• BTHA uses shape-preserving adapters to inject semantic guidance while maintaining a stable feature-level interface across different backbones.
• The method introduces a Hierarchical Coarse-to-Fine Supervision Strategy for global alignment, multi-scale localization, and boundary refinement.
• A Scale-Adaptive Gated Semantic Guidance (SAGSG) adapter adaptively controls textual injection and suppresses redundant cross-modal responses.
• Evaluations show the framework remains effective across convolutional and transformer-based visual encoders as well as various language encoders.

Experiments on four public datasets demonstrate that BTHA improves strong text-guided baselines with modest computational overhead.