Currently, most deep learning models are still black-box methods. Particularly in medical applications, not only the performance but also the explainability of artificial intelligence (XAI) is crucial. In this paper, models are examined which are used to diagnose patients with ischemic strokes. Two different methods of XAI are used to explain deep neural networks: Grad-Cam and Grad-CAM++. They are used to show where the class discriminating pixels are located on a Magnetic Resonance (MR) image. Various experiments show how the explainability of deep learning models can be improved and whether general statements can be made for the classification of stroke patients based on their MRIs. Furthermore, a convolutional neural network (CNN) is implemented to detect an existing stroke and classify the severity of the disability it causes (mRS Outcome). With an accuracy of 94%, images of patients can be classified as stroke or no stroke. The XAI shows that both models can reliably detect brain lesions caused by a stroke. Although the data is unbalanced, the model that predicts the mRS outcome has an overall accuracy of 92%. It is shown that there are differences in the explanations for mRS Outcome 3-6 and mRS Outcome 0-2. By using Grad-CAM, lesions in the brain can be detected, which are even overseen by experienced neurologists. In addition, it is possible to simplify models without significant performance loss by using Grad-CAM. The resulting explanations can thus serve experts such as neurologists and physicians as a basis for new hypotheses or even help them to improve their diagnostic quality.