Whole slide imaging (WSI) plays a pivotal role in advancing disease diagnosis, particularly in pathology, by providing high-resolution digital representations that allow for detailed analysis of tissue structures in diseases such as cancer. Multiple-instance learning (MIL), as a weakly supervised learning approach, has demonstrated excellent performance in WSI analysis in recent years, especially in cancer classification and detection. However, a significant challenge in current research is that MIL methods tend to focus on highly discriminative instances, resulting in overfitting. To address this issue, we propose a novel approach called a multi-branch independent masking and Dirichlet-based fusion method for multiple instance learning on whole-slide images (MD-MIL). Specifically, we first introduce the multi-branch independent masking (MIM) component, which aims to obtain multi-branch prediction results by masking different instances for different branch. This prevents overemphasis on prominent features such as cell nuclei, ensuring the model can flexibly attend to a broader range of regions. Secondly, we utilize the Dirichlet-based multi-branch fusion (DMF) component, which leverages uncertainty estimation to fuse predictions from multiple branches, thereby enhancing classification accuracy. To evaluate the performance of this approach, we conducted extensive experiments on the CAMELYON-16, BRACS, and TCGA-LUNG datasets, with results demonstrating that MD-MIL outperforms other state-of-the-art methods.