Hypoxic ischemic encephalopathy (HIE) affects approximately 1 to 5 per 1000 live births, with nearly 20% of cases proving fatal. This condition arises when a neonate’s brain receives insufficient oxygen or blood flow during or shortly after birth. Even among survivors, there is a high risk of long-term neurocognitive and behavioral disorders, sensory impairments, and epilepsy. Timely diagnosis and accurate segmentation of HIE-affected regions are critical for initiating appropriate medical intervention, and magnetic resonance imaging (MRI) is a widely utilized tool for this purpose. In this study, we addressed the diagnosis and segmentation of HIE using a 3D-MRI dataset from BONBID-HIE 2024 Grand Challenge. A significant challenge posed by this dataset is the small lesion size typically <1% of total brain volume and their diffused nature. To address class imbalance, we applied dynamic class weighting into the loss function, adjusting the weights for each batch based on class distribution. For the segmentation task, we developed a novel heterogeneous ensemble approach, combining the strengths of three state-of-the-art 3D deep learning models: SegResNet, Swin-UNETR, and UNet. This ensemble strategy was designed to leverage the complementary capabilities of each architecture to improve performance on the complex lesion segmentation problem and was able to achieve the dice score of 57.7±23.7 on test leader board.