Hybridization of Attention UNet with Repeated Atrous Spatial Pyramid Pooling for Improved Brain Tumor Segmentation

Brain tumors are highly heterogeneous in both spatial distribution and scale, making their segmentation in medical images a challenging task that can lead to diagnostic and therapeutic errors. Automating tumor segmentation has the potential to enhance objectivity and repeatability while significantly reducing turnaround time. Conventional convolutional neural networks (CNNs) often struggle to accurately capture the wide range of tumor sizes and shapes, leading to suboptimal performance. To address these limitations, UNet has emerged as a widely adopted solution for semantic segmentation, leveraging a downsampling-upsampling architecture to segment tumors. Building on this foundation, we propose a novel architecture that combines Attention-UNet with repeated Atrous Spatial Pyramid Pooling (ASPP). ASPP captures multi-scale contextual information effectively through parallel atrous convolutions with varying dilation rates. Our approach demonstrates substantial improvements over UNet, Attention UNet, and Attention UNet with Spatial Pyramid Pooling, establishing a new benchmark for tumor segmentation tasks.