SLAM-FusionNet: A Deep Learning Framework Integrating Spatial Local Attention with Multi-Modal Fusion for Glioma Segmentation

The accurate segmentation of gliomas and subregions of multimodal magnetic resonance imaging (MRI) plays a significant role in the accurate diagnosis of the disease and formulation of therapy as well as disease surveillance. However, the non-uniformity of tumours where boundaries vary and modality variations are specific to the intensity subjects a person to a huge challenge. The work will present SLAM-FusionNet a transformer-based architecture, a Multi-Modal Fusion (MMF) strategy and a Spatial Local Attention Module (SLAM) in order to address effectively the local and global contextual information of tumor fine-grained information. MMF module can be incorporated to increase cross-modality representation learning, complementary characteristics of T1, T2, FLAIR and T1ce images are merged; to increase localization in a spatial dimension, SLAM increases the significance of the spatially-relevant and boundary-sensitive regions that can better differentiate intra-tumor subregions. The suggested network is premised on a Swin Transformer backbone, and the primary strength of long-range dependency description and local spatial fidelity. Considerable testing has been carried out on the BraTS dataset that demonstrates that SLAM-FusionNet is functional with 95.6% whole tumor (WT) 96.2% tumor core (TC) and 94.8% enhancing tumor (ET) Dice scores and an average Dice of 95.5. The average HD95 is also increased to 3.95 mm and improved compared to state of the art models such as Swin-UNet and nnU-Net. Additive value of MMF and SLAM is confirmed by formal studies of ablation. The results highlight the applicability and clinical power of SLAM-FusionNet in computer-aided brain tumor segmentation in precision neuro-oncology.