Hyperspectral pathology images provide rich spectral and spatial information, enabling precise tissue analysis. However, their high dimensionality and complex structures pose significant challenges for accurate segmentation. To address this issue, we propose a Dual-Track Graph Fusion Network (DTGFN). First, to reduce data redundancy, the first principal component of the raw data is processed with a Gaussian kernel convolution, ensuring that superpixel regions preserve more accurate tissue boundaries. Subsequently, we constructed a dual-branch network, in which the graph convolutional networks (GCN) branch focuses on capturing local neighborhood features, while the hypergraph convolutional network (HGCN) branch models cross-regional high-order dependencies, enabling a comprehensive representation of both local and global information. Channel attention mechanisms are incorporated in each branch to adaptively enhance key channels and improve feature discriminability. Finally, experiments on a cholangiocarcinoma hyperspectral pathology dataset demonstrate that DTGFN significantly outperforms existing methods in Dice and IoU metrics, while providing more precise delineation of lesion boundaries, offering a robust and generalizable solution for high-precision hyperspectral pathology image segment.