Hypoxic-ischemic encephalopathy (HIE) significantly impacts neurological development in infants, and accurate lesion segmentation from MRI images is crucial for diagnosis and treatment. However, traditional deep learning models often struggle with HIE’s diverse lesion characteristics. This paper presents a novel ensemble strategy utilizing Swin-UNETR, a transformer-based model, to address this challenge. We demonstrate the advantages of Swin-UNETR for HIE lesion segmentation compared to U-Net 3D, a commonly used model. By leveraging self-attention modules and hierarchical encoding, Swin-UNETR captures long-range dependencies and contextual information, leading to superior performance across all evaluation metrics (Dice, MASD, NSD). Our ensemble, combining five Swin-UNETR models trained on different folds, further improves these results, ranking within the top 3 methodologies in the BONBID-HIE challenge.