Moderate to severe traumatic brain injury (msTBI) results in complex and heterogeneous brain lesions, posing significant challenges for neuroimaging analysis. These lesions vary widely in size, number, and tissue distribution, complicating processes such as image registration and parcellation. Traditional segmentation tools often require multiple imaging modalities or manual intervention, limiting their effectiveness for msTBI. To address these challenges, we propose NeuroNetMix, a 3D encoder-decoder framework specifically designed for T1-weighted MRI, the most widely available scan type in the ENIGMA TBI consortium. NeuroNetMix integrates bidirectional quasiseparable mixing at its bottleneck to capture both local lesion detail and long-range spatial dependencies, enhancing volumetric continuity and segmentation accuracy. The encoder extracts hierarchical volumetric features while preserving small and irregular lesions via skip connections, and the decoder reconstructs high-resolution segmentation maps. Evaluations demonstrate that NeuroNetMix outperforms existing methods, including xLSTM and 3D Mamba UNet, achieving state-of-the-art performance on both validation and testing datasets. By providing reliable and precise lesion maps, NeuroNetMix facilitates improved downstream analyses such as brain parcellation and connectomics, supporting more accurate prognostic assessment and treatment planning for msTBI patients.