In computational pathology, the method of multiple instance learning (MIL) has become a common solution for processing such high resolution images. However, most previous studies have focused on extracting features at a single scale for analysis, neglecting the multi-scale information crucial for high-resolution images. The few existing studies based on multi-scale features cannot integrate cross-scale features well, which is mainly reflected in the fact that the model over-relies on high-scale features when identifying simple samples, while ignoring the learning of low-scale features, resulting in poor performance of the model when mining difficult samples and poor generalization ability of the model. To address the above problems, we designed a novel cross-scale MIL model for pathological image diagnosis. We introduced the improved hard instance mining strategy into multi-scale features, aiming to balance the contribution of each scale feature to the classification result and avoid the excessive dominance of high-scale features in classification decisions. Additionally, we proposed a novel cross-scale cross-attention mechanism to achieve full interaction of features of different scales. Extensive experiments demonstrated the effectiveness of our model, which outperformed other advanced methods on the TCGA lung cancer public dataset.