Precise segmentation of gastric cancer computed tomography (CT) images is a critical step for clinical precision diagnosis and treatment. However, it currently faces two major challenges: the low contrast between tumors and surrounding normal tissues makes boundary delineation difficult, and the high variability in tumor shape, size, and location leads to inaccurate localization. To address these issues, a cross-modal prior knowledge-guided gastric cancer CT image automatic segmentation method (CGP-Net) was proposed. In this method, visual priors were extracted from diagnostic reports using a large language model (LLM), and lesion localization was assisted by a semantic anchoring and parsing module. A mixed context-aware Mamba module was constructed to synergistically optimize feature modeling for adapting to tumor morphological variations. Furthermore, a boundary-aware gated convolution module was designed to improve the delineation accuracy of fuzzy boundaries. Experiments on a large-scale dataset of 349 gastric cancer patients demonstrated that the Dice coefficient and 95th percentile of Hausdorff distance (HD95) of the proposed method reached 78.10% and 16.44 mm, respectively. It outperformed state-of-the-art methods such as U-Mamba and nnUNet in terms of segmentation accuracy and boundary prediction. This method effectively integrates textual priors to significantly enhance segmentation accuracy, offering significant value for clinical applications.
