Microvascular obstruction (MVO) is an important imaging biomarker associated with adverse prognosis following reperfusion therapy in acute myocardial infarction. In late gadolinium enhancement cardiac magnetic resonance (LGE-CMR) images, MVO typically appears as small hypo-intense regions embedded within scar tissue. Owing to its small size, ill-defined boundaries, and severe class imbalance, accurate and robust automatic segmentation of MVO remains challenging for existing deep learning methods. To address these challenges, this study proposes a hierarchical segmentation framework based on regional prior constraints and adaptive bias correction. Firstly, a three-stage cascaded segmentation model is constructed to progressively segment the myocardial ring, scar region, and MVO, where anatomical spatial priors are incorporated to gradually constrain the search space. Secondly, a bias-aware mechanism is introduced to identify difficult pixels, and cross-sample similar pixel replacement is employed to achieve adaptive enhancement of hard regions, thereby alleviating model bias during training. In addition, a structurally constrained loss function based on contour distance distribution modeling is proposed to improve spatial consistency from a global shape perspective. Based on the public EMIDEC dataset, extensive experimental results demonstrate that the proposed method achieves significant improvements over several state-of-the-art approaches in terms of Dice coefficient and 95% Hausdorff distance (HD95), validating its effectiveness and robustness for MVO segmentation.