The precise, automated segmentation of lesions in moderate-to-severe traumatic brain injury (msTBI) is a formidable clinical challenge, hindered by their minute size, erratic shapes, and diverse appearances. This paper introduces DeaMNet, a novel dual-encoder network with a hybrid morphology-aware loss, designed for the MICCAI 2025 AIMS-TBI Challenge. The novelty of DeaMNet lies in its synergistic combination of three key innovations: (1) a dual-encoder architecture that simultaneously extracts deep semantic features and lightweight spatial details, (2) multi-level skip connections that effectively fuse hierarchical features into the decoder, and (3) a hybrid morphology-aware loss function that addresses severe class imbalance and boundary ambiguity. DeaMNet leverages a powerful ResNet backbone in parallel with a lightweight MobileNetV2 backbone to comprehensively learn both the complex high-level patterns and the fine-grained boundary characteristics of TBI lesions. The proposed hybrid loss, which combines Dice, contrast-weighted Binary Cross-Entropy (BCE), and Boundary Loss, guides the model to focus on irregular and ambiguous lesion boundaries. DeaMNet significantly outperforms the standard DeepLabV3+ baseline and other common segmentation models, demonstrating this synergistic approach is a key strategy for maximizing performance in complex TBI MRI analysis.