Atrous Convolution-Based Fusion Attention Mechanism for Brain Tumor Segmentation

Accurate medical image segmentation is pivotal for advanced diagnostic and therapeutic planning, especially for intricate tasks such as brain tumor delineation. However, existing segmentation methods often struggle with challenges such as heterogeneous tumor morphologies, diffuse boundaries, and low-contrast regions, limiting their reliability in complex medical imaging scenarios. In this work, we introduce an Atrous Convolution-Based Fusion Attention Mechanism, a novel framework that combines local and global attention through an innovative fusion block. Local attention leverages atrous convolutions to capture fine-grained spatial features at multiple scales, while global attention extracts long-range contextual dependencies across the image. These complementary features are dynamically fused via a novel fusion block, ensuring balanced representation and robust feature aggregation for precise segmentation. The proposed model is rigorously evaluated on the BraTS 2019, 2020, and 2023 datasets, demonstrating consistent and significant improvements across key segmentation metrics compared to baseline models. Quantitatively, the model achieves an absolute average increase of 0.2522, 0.1873, and 0.1934 in the Dice Coefficient, and 0.3071, 0.2365, and 0.2369 in the Intersection over Union (IoU) for enhancing tumor, tumor core, and whole tumor, respectively, across all datasets. The results highlight robustness and adaptability of the proposed architecture, underlining its potential to address critical challenges in brain tumor segmentation and advance the state-of-the-art in medical image analysis.