Multi-Instance Learning (MIL) treats whole slide images (WSIs) as “bags” and the unlabeled patches as “instances”, enabling weakly supervised classification via attention-based feature aggregation. However, conventional attention mechanisms may fail to capture diverse pathological patterns, relying on a few highly discriminative instances while overlooking potential patterns. Additionally, spatial continuity in WSIs is disrupted during instance partitioning, limiting the ability of the model to perceive global lesion distribution. To address these challenges, we propose SAM-Enhanced Multi-Branch Attention Multi-Instance Learning (SMBA-MIL). We introduce an Adaptive Multi-Branch Attention Aggregation (AMBA) module with parallel attention branches to capture heterogeneous pathological patterns, and a Dynamic Branch Weight Generator (DBWG) to adaptively fuse multi-branch semantics. Moreover, we employ the Segment Anything Model (SAM) to segment WSIs and construct positive and negative instance pairs based on whether they belong to the same segmentation region. Contrastive learning is applied at the attention distribution level, leveraging the global spatial information contained in segmentation regions to guide the multi-branch attention toward better perception of the overall lesion distribution. Extensive experiments on the Camelyon and TCGA-Lung datasets show that SMBA-MIL achieves superior performance in terms of accuracy and AUC, demonstrating its ability to combine both pathological pattern diversity and spatial awareness.