In this paper, we propose an improved U-Mamba framework based on domain generalization techniques. The framework is capable of automatically segmenting normal myocardium, edema, and scar from multi-sequence and multicenter cardiac magnetic resonance images. In the data preprocessing stage, we addressed the issue of domain shift in multicenter data by histogram matching. Conventional segmentation strategies often overlook the intercorrelation of myocardial pathologies (scar and edema). To solve this, we innovatively introduced a region-based training strategy which focuses on the overall segmentation of target regions, which is more in line with theoretical medical knowledge. In the validation set of the MyoPS++ challenge, our proposed framework achieved mean Dice scores of 0.6575 and 0.6855 in LV myocardial scar and the union of scar and edema, respectively. Our code is available at https://github.com/gaojh135/MyoPS2024 .