Fine-grained segmentation of spine MRI into vertebrae, intervertebral discs, and spinal canal, is critical for diagnosing complex spinal disorders. However, existing methods, which primarily rely on visual features, struggle with capturing the global anatomical understanding and pixel-level semantics. To address these challenges, we propose DP-Net, a dual-prompt-enhanced visual-language model with two key modules: The text insight enhancement module at the omni-level (TIEO) integrates global image context and anatomical distribution awareness into visual features, enhancing the model’s global understanding of the image. While the text insight enhancement module at the pixel-level (TIEP) further refines segmentation details by aligning semantics at the pixel level. These prompts are adaptively generated from the existing training data without additional textual annotations. Extensive experiments on the SPIDER and MRSpineSeg datasets demonstrate that DP-Net outperforms existing promising methods. For example, DP-Net achieves DSC \(\uparrow \) of 84.58 and HD95 \(\downarrow \) of 3.56 on SPIDER, surpassing the strong baseline model Swin-UMamba by +4.47% and −2.68, respectively.