| Summary: | This study introduces HMSA-Net, a Hierarchical Multi-Scale Attention Network designed for accurate brain tumor segmentation from multi-modal MRI scans. HMSA-Net employs a hierarchical encoder to capture multi-resolution features, complemented by full-scale skip connections that preserve spatial detail across different scales. In the decoder, dynamic multi-scale attention modules recalibrate and fuse features, enhancing the network’s ability to differentiate complex tumor subregions. The proposed method is evaluated on the BraTS 2021 dataset and demonstrated consistent improvements over baseline models in segmentation accuracy and boundary localization. Additionally, a bagging-based ensemble strategy improves prediction stability and enhances robustness in the presence of tumor heterogeneity. By integrating hierarchical multi-scale feature extraction with attention-driven refinement, HMSA-Net provides a reliable and efficient solution for automated brain tumor segmentation, with significant potential to support quantitative neuro-oncological analysis and inform clinical decision-making.
|
|---|
