<p>Accurate lesion segmentation is essential for diagnosing and treating strokes. However, the task of segmenting stroke lesions is difficult due to the intricate backgrounds and noise prevalent in medical images. To tackle this challenge, the proposed research introduces an enhanced U-Net architecture integrating pre-trained CNN blocks, bottleneck residuals, and an adaptive gated neuro transformer with a modified MLP. This novel architecture is structured to assimilate both local and global aspects, thereby improving segmentation accuracy. The pre-trained CNN blocks in the architecture take advantage of features learned through supervised tasks, which enhances the model's generalization abilities. Bottleneck residual blocks are incorporated to perform high-level feature extraction and dimensionality reduction, that is essential for managing the complexities inherent in medical imagery. The adaptive gated neuro transformer, equipped with a novel MLP, will efficiently captures long-range correlations by utilizing spatial data, further refining the segmentation process. To confirm our model's efficacy, we conducted trials with the ATLAS and ISLES2022 collections. Observations showed that our model markedly surpassed traditional approaches, highlighted by substantial enhancements in the Dice Similarity Coefficient (DSC) and decreases in the Hausdorff Distance (HD). Such indicators are vital for assessing the precision and reliability of segmentation methods. In addition to its quantitative success, the proposed method excels qualitatively by preserving edges and accurately detecting lesion boundaries. This indicates that the proposed improved U-Net transformer architecture is not only effective in segmenting stroke lesions but also holds considerable promise for clinical application. By enhancing the precision of stroke lesion segmentation, the proposed model is potentially contributing to better diagnosis and treatment planning, ultimately enhancing patient care.</p>