Early and accurate diagnosis of ischemic stroke is essential to improve patient treatment and prognosis. In this context, automatic lesion segmentation in DWI images constitutes a valuable tool to assist specialists in identifying and quantifying affected areas. In this work, we propose an approach based on deep learning using an Attention U-Net model, designed to enhance lesion localization by utilizing attention mechanisms. This approach achieves a DICE Similarity Coefficient (DSC) of 0.78 in acute phase images (<6 h post-infarction) and demonstrates generalization in follow-up images (48 h post-infarction). The model was trained and validated with data from 55 patients, employing data augmentation and cross-validation to address the limited dataset size. The results showed high sensitivity (0.85) and an F1 score of 0.89 with an optimal threshold of 0.90, indicating that this approach may be useful in supporting the diagnosis of cerebral infarction.