Enhanced Brain Tumor Segmentation in MRI using SwinSpatialNetwork: Integrating SLAM and ELSA for Superior Multiscale Context Analysis

Brain tumor segmentation from MRIs is crucial for tracking tumor growth and planning surgical interventions. Current methods face challenges in precisely specifying border features and integrating global context data. We introduce SwinSpatialNetwork, an advanced architecture that combines CNNs with Swin Transformers for precise brain tumor extraction. The model includes two key modules: the Spatial Local Attention Module (SLAM) and the Enhanced Local Self-Attention (ELSA) transformer block. SLAM focuses on regions of interest, improving the network's capacity to identify subtle differences, while ELSA facilitates detailed local feature extraction. Evaluation using the BraTS 2021 dataset demonstrates exceptional segmentation performance, with our method achieving an average Dice Similarity Coefficient (DSC) of 89.75% and a 95% Hausdorff Distance (HSD) of 3.20 mm. These results surpass the performance of advanced 3D algorithms. SwinSpatialNetwork marks a significant breakthrough in brain tumor segmentation by effectively addressing the challenges of lesion variation and minor visual similarity. This model offers an accurate and reliable solution for medical imaging segmentation, setting a new standard in the field.