Automated Segmentation of Acute Ischemic Stroke Using Attention U-Net with Patch Mechanism

paper addresses ischemic stroke detection using deep learning techniques to interpret medical images like MRI and CT scans, with a focus on segmentation. Ischemic stroke occurs when a blockage in brain arteries disrupts blood flow, impairing brain functions. The study aims to develop a model for automatic segmentation of ischemic stroke areas, facilitating efficient diagnosis in medical settings. An enhanced Attention U-Net model with a patch-based approach using MRI data is proposed for this purpose. The model was validated on the ISLES'22 public ischemic stroke dataset. The segmentation process consisted of three stages. First, the standard Attention U-Net model achieved a Dice Similarity Coefficient (DSC) of 88.9%. In the second stage, the MRI images were divided into 32x32 patches and reanalyzed, increasing the DSC to 93%. In the final stage, different attention mechanism methods were added to the U-Net architecture and the effect of attention mechanism on segmentation success was observed. Asa result of the experiments, the U-Net architecture using spatial attention achieved 94.86%, Athe U-Net architecture using SE attention achieved 95.40%, Aand the U-Net architecture using CBAM attention achieved 96.47% DCS success. The study concludes that the enhanced model outperforms existing methods, demonstrating that the proposed approach is effective for segmenting ischemic strokes and yielding significant results compared to similar studies in the literature.