Diversity-driven MG-MAE: Multi-granularity representation learning for non-salient object segmentation | Synapse