ASPP-HAM-UNet: a modified UNet for moderate-to-severe traumatic brain injury segmentation

Traumatic brain injury (TBI) lesion segmentation from magnetic resonance imaging (MRI) remains challenging because lesions exhibit substantial heterogeneity in size, shape, intensity, and anatomical distribution. Conventional UNet-based segmentation models often struggle to accurately delineate fragmented or low-contrast lesion boundaries while maintaining contextual consistency. In this study, we propose Atrous Spatial Pyramid Pooling (ASPP)-Hybrid Attention Module (HAM)-UNet, a modified UNet architecture that integrates ASPP and a HAM for moderate-to-severe TBI (msTBI) lesion segmentation in T1-weighted MRI. The ASPP module is incorporated at the bottleneck layer to capture multi-scale contextual information, while HAM is utilized within gated skip connections and decoder refinement stages to enhance channel-wise and spatial feature representation. The proposed framework was evaluated on the publicly available AIMS-TBI2025 dataset using the Dice coefficient (DC), precision, recall, and 95th percentile Hausdorff Distance (HD95). Experimental results demonstrate that ASPP-HAM-UNet achieved a DC of 0.8747 and an HD95 value of 8.9937, outperforming standard UNet and Attention UNet while maintaining moderate computational complexity. Extensive ablation studies further demonstrate the synergistic contribution of ASPP and HAM in improving lesion localization and boundary delineation. These findings indicate that the proposed ASPP-HAM-UNet provides an effective framework for automated segmentation of msTBI lesions in MRI.