Multi-instance multi-label learning (MIML) provides a framework for handling complex objects represented by multiple instances and associated with multiple labels. However, existing methods often overlook intrinsic correlations among instances and lack transparency in decision-making, which limits their accuracy and interpretability. To address these issues, we propose an advanced MIML framework based on a Gated Graph Attention Network (MIML-GGAT). Our approach represents instances as nodes in a graph to explicitly display their relationships and employs Graph Convolutional Networks (GCNs) to enrich node features with contextual semantics. Furthermore, a gated attention mechanism is introduced to generate label-specific weights, enabling the model to automatically focus on critical instances for each label and providing clear decision evidence. Experimental results on Breast cancer semantic segmentation dataset demonstrates that MIML-GGAT achieves competitive and superior performance compared to other methods. Simultaneously, it has significantly enhanced interpretability, thereby validating its effectiveness in medical image diagnosis applications.