Accurate segmentation of 3D colon meshes is pivotal for virtual colonoscopy but is hindered by the challenge of capturing both local geometric details and global anatomical context. Existing mesh deep learning methods, often limited by local receptive fields, struggle to model the long-range dependencies inherent in the colon’s serpentine structure. To address this, we introduce ColonMeshNet, a novel hierarchical deep learning architecture that synergistically fuses local and global feature learning. Its core innovation is a ConvEdgeAttention module, which integrates a lightweight convolutional operation directly into a self-attention mechanism. This design effectively injects a local inductive bias, enabling the model to learn robust, context-aware representations at multiple scales. To ensure a rigorous and standardized evaluation, we constructed a dataset of 100 expertly annotated clinical colon meshes. Furthermore, we propose a complete processing pipeline with a robust backward mapping framework to transfer predictions from simplified to high-resolution meshes. Experiments demonstrate that ColonMeshNet significantly outperforms state-of-the-art methods, achieving 80.2% accuracy on our evaluation set.