Gba-Net: A Gated Bottleneck and Attention-Driven Architecture for Robust Ischemic Stroke Segmentation across Ct and Dwi

Abstract

Automated ischemic stroke segmentation remains difficult because non-contrast Computed Tomography (CT) is low contrast and noisy, whereas Diffusion Weighted Imaging (DWI) shows heterogeneous lesions. Conventional U-Net models rely on local receptive fields and unselective skip fusion, limiting global context and noise control. We propose GBA-Net, a UNet variant that combines a high-capacity gated Convolutional Neural Network (CNN) bottleneck for long range dependencies with convolutional block attention modules that refine multiscale features before decoder fusion. The bottleneck helps interpret subtle CT hypo densities and link scattered infarcts in DWI, while attention suppresses CT noise and filters high intensity mimics, improving boundary delineation. We evaluated GBA-Net on ISLES 2024 and TEKNOFEST 2021 and compared it with nine baselines including UNet, UNet++, DeepLabV3+, and Seg-Former. GBA-Net achieved Dice 0.7376 and 0.7140 and the best average ASSD of 4.73 pixels on CT.