Curvilinear structure segmentation in diverse real-world scenarios requires flexible interactive models that can accurately interpret user intent. While the Segment Anything Model (SAM) excels in visual segmentation, its unimodal reliance on visual prompts limits its adaptability for such tasks. We propose a novel multimodal framework that enhances curvilinear structure segmentation by unifying vision and language. Our approach replaces SAM’s visual prompt encoder with a pre-trained CLIP text encoder, enabling precise language-guided segmentation through natural language instructions. To improve sensitivity to curvilinear patterns, we introduce a wavelet-based structure-aware enhancement module that injects high-frequency residual information into each image encoder block, complementing spatial features with frequency-domain insights. Additionally, we incorporate adapter layers at specific encoder blocks to facilitate efficient domain adaptation with minimal parameter overhead. Extensive evaluations across diverse domains demonstrate that our framework achieves superior segmentation performance while excelling in responsive, language-driven segmentation. This work sets a new benchmark for interactive multimodal segmentation of curvilinear structures.