This study focuses on the automated segmentation of acute ischemic stroke lesions from non-contrast CT (NCCT) scans using deep learning. Researchers have developed a novel 3D deep learning model called CoAt U SegNet, which is able to integrate convolutional and attention mechanisms together. The model was trained on 60 NCCT scans and tested on 500 real-world clinical scans to evaluate its performance. The proposed model achieved a high Dice similarity coefficient of 75.7% and a Jaccard index of 70.8%, outperforming other standard models such as 3D U-Net, IS-Net, and EIS-Net. The new model was also particularly effective in identifying lesions near low-contrast areas, such as the lateral ventricles, and showed strong statistical significance with a p-value of 0.0008 and a large Cohen’s d effect size of 1.72. Both visual and statistical results confirmed the model’s robustness across a diverse set of clinical cases. Lastly, the discussion highlights the model's superiority in challenging the NCCT scenario and its hybrid architecture as a key strength. Additionally, researchers acknowledge limitations such as the need for multi-center validation and distinguishing between the infarct core and penumbra. Future works and evaluations are said to focus on improving generalizability, clinical deployment, and integration with outcome-based tools.