Automated segmentation of intracranial hemorrhage (ICH) is an important clinical task limited by the scarcity of high-quality annotated data and variability across datasets. To address these challenges, we propose a novel two-phase learning framework that leverages self-supervised pretraining and multi-label decoding. First, we pretrain a hierarchical vision transformer on over 25,000 head CT cases using a Masked Autoencoder (MAE) strategy to learn robust multi-scale feature representations. For the downstream task, we append a multi-label decoder with dataset-specific heads, which enables joint learning from four distinct ICH segmentation datasets with diverse annotation protocols. A tailored sampling strategy further addresses class imbalance and heterogeneity across datasets. Our results demonstrate superior performance, with the proposed model achieving a mean Dice score of 63.57% (±12.30), surpassing the next best-performing method by 2.69%. Notably, the proposed MAE-based model outperformed its BYOL-pretrained counterpart, suggesting that MAE’s reconstruction-based objective is a more effective pretraining strategy for learning the dense, spatially-rich representations required for this segmentation task.