Purpose <p>This study aims to develop an autonomous segmentation approach for acute ischemic stroke lesions in diffusion-weighted imaging using deep learning techniques in the acute phase. The main research revolves around the efficacy of an introduced specialized deep-supervised U-Net architecture in accurately segmenting these lesions from diffusion-weighted imaging images.</p> Methods <p>The proposed method employs a deep supervised U-Net architecture with specialized convolution blocks comprising six parallel layers. The model is trained using the ischemic stroke lesion segmentation challenge 2015 Dataset, and extensive data pre-processing and augmentation techniques are utilized. The model’s performance is evaluated using quantitative metrics, including the dice similarity coefficient, on the test dataset.</p> Results <p>The proposed autonomous acute ischemic stroke lesion segmentation approach achieves impressive results, with an average dice similarity coefficient of 0.78 on the test dataset. Comparative analysis with conventional segmentation methods reveals superior performance, especially in handling challenges such as intensity inhomogeneity and image artifacts. Notably, the model exhibits robustness and generalizability across diverse image datasets.</p> Conclusion <p>The research demonstrates that the introduced deep learning technique accurately identifies acute ischemic stroke lesions, overcoming manual measurement issues and typical segmentation challenges like intensity irregularities, and artifacts, and also improves lesion detection accuracy. These findings contribute to advancing stroke diagnosis and treatment planning, ultimately improving healthcare outcomes for stroke patients.</p>