For doctors to focus on certain parts of the brain and provide patients the appropriate therapy, accurate brain stroke diagnosis, categorization, and segmentation are crucial. Many artificial intelligence applications have successfully included encoder-decoder deep learning techniques. However, because of their effective operations, efficient sampling techniques, and learning procedures, such networks have several drawbacks. In this work, U-Net, a deep learning-based encoder-decoder convolutional neural network (CNN), has been developed and proposed for the segmentation and categorization of brain strokes. The suggested method examines the images from the computed tomography (CT) dataset that was used to detect the presence of a brain stroke. Once a stroke has occurred, it is possible to ascertain whether it was caused by ischemia. The suggested approach can also precisely segment the existing stroke and highlight the area that the radiologist has overlay. By using Python scripts to run a number of trials on the same actual dataset, the suggested technique is contrasted with alternative CNN-type architectures that are already in use. The precision of classification rates for stroke at 99.2% & ischemia with hemorrhage at 99.3%, respectively, demonstrate the efficiency of the proposed methodology. Additionally, the suggested model's intersection over union (IoU) rate for segmenting brain strokes was 95.2%.