Rodent models are essentials for studying the mechanisms underlying traumatic brain injury (TBI) and its long-term effects. Magnetic resonance imaging (MRI) is key to monitor in-vivo structural changes, but accurate segmentation of brain volumes remains a significant challenge. This is particularly evident in the case of focal contusion where anatomical distortions reduce the accuracy of registration-based methods, often necessitating time-consuming manual corrections. To overcome these limitations, we developed a Convolutional Neural Networks (CNNs)-based tool to enhance segmentation quality and efficiency. Using a dataset of mice and rats subjected to TBI by controlled cortical impact (CCI), we evaluated the performance of a 3D multi-task CNN across mouse strains, rodent species, injury sites and MRI sequences. By employing multi-scale, deep supervision and attention mechanisms with a five-fold cross-validation approach, we achieved an average Dice score above 0.98 for skull-stripping and over 0.84 for lesion segmentation. We then tested a domain adaptation strategy to handle mixed partial annotations and transfer features across both injured and healthy rodent subjects, enabling automatic segmentation of the cortex, hippocampus, corpus callosum and ventricle, in both hemispheres. This approach achieved a Dice score of 0.98 for skull-stripping and a mean Dice score of 0.88 for the ten classes. Finally, we validated the CNNs lesion segmentation by comparing its results with those of a trained operator, finding a strong correlation (Pearson r = 0.974) and a 19% reduction in variability (SD decreased from 3.29 to 2.66). This robust generalization could help standardize analyses, enhance across-laboratories comparability, and improve clinical translational potential.