Pathological images play an irreplaceable role in cancer diagnosis; however, their ultra-high resolution and complex structures pose significant challenges for automated analysis. Although deep learning has made remarkable progress in recent years, clinical practice still heavily relies on manual expert annotation when dealing with large-scale, multi-resolution pathology images. In particular, due to the enormous size and high annotation cost of whole-slide images (WSIs), it is often infeasible for pathologists to perform exhaustive annotations, making it crucial to explore how sparse annotations can effectively guide full-slide automatic segmentation. To address this issue, we propose a region graph-driven weakly supervised segmentation framework that integrates structure-aware modeling with confidence-guided optimization to achieve efficient and scalable lesion recognition. Specifically, the method first performs fine-grained over-segmentation on downsampled WSIs to construct region-level graphs, and then leverages graph convolutional networks (GCNs) along with cross-region attention mechanisms to aggregate global contextual information. For annotated regions, the model is trained using cross-entropy loss, while for unannotated regions, a confidence estimation module is introduced to generate high-quality pseudo-labels for further unsupervised refinement. The region graph structure ensures label consistency among visually homogeneous areas and enhances semantic representation and classification by modeling long-range dependencies between graph nodes. Experiments on real-world cancer pathology datasets demonstrate that the proposed method can accurately localize and segment lesion regions using only 2% of annotated data, offering a promising new paradigm for computational pathology.