RAX-NET: Residual Attention Xception Network for Brain Ischemic Stroke Segmentation in T1-Weighted MRI

Ischemic stroke, caused by arterial occlusion, leads to hypoxia and cellular necrosis. Rapid and accurate delineation of ischemic lesions is essential for treatment planning but remains challenging due to variations in lesion size, shape, and appearance. We propose Residual Attention Xception Network, a deep learning architecture that integrates residual attention connections with Xception for three-dimensional magnetic resonance imaging lesion segmentation. The framework includes three stages: (i) decomposition of three-dimensional scans into axial, sagittal, and coronal planes, (ii) independent model training on each plane, and (iii) voxel-wise majority voting to generate the final three-dimensional segmentation. In addition, we introduce a variant of the focal Tversky loss designed to mitigate class imbalance and improve sensitivity to small or irregular lesion boundaries. Experiments on the ATLAS v2.0 dataset with five-fold cross-validation demonstrate that Residual Attention Xception Network achieves a Dice coefficient of 0.61, precision of 0.68, and recall of 0.63. These results surpass baseline models while requiring fewer trainable parameters and enabling faster inference, highlighting both accuracy and efficiency.