Lesion segmentation in Moderate to Severe Traumatic Brain Injury (msTBI) is challenging due to the wide variability in lesion size, location, and appearance. The AIMS-TBI Segmentation Challenge 2025 provided a benchmark for evaluating methods on T1-weighted MRI data. In our submission, we focus on a deliberately simple approach: training standard U-Net models without pretraining or external datasets, and exploring how postprocessing and ensembling strategies can improve performance. We find that an ensemble of just two models yields consistent gains over single models, and that performance can be further improved by applying a two-stage ensembling scheme: first at the level of lesion detection (presence vs. absence), and then at the level of segmentation map aggregation. Our results demonstrate that high-quality msTBI lesion segmentation does not necessarily require large-scale pretraining or complex networks. Instead, careful ensemble design and simple postprocessing are effective levers for boosting performance in the challenge setting.