Stroke is a group of diseases with severe brain tissue damage, which are caused by either the sudden rupture of brain blood vessels (cerebral hemorrhage) or brain blood vessel obstruction, leading to rapid changes and high mortality. The diagnosis of stroke mainly relies on medical imaging techniques, including Computed Tomography (CT) and Magnetic Resonance Imaging (MRI), which require experienced radiologists to guarantee suitable accuracy. However, the amount of brain CT image data is extremely large, usually exceeding the technical capabilities of radiologists. Currently, artificial intelligence has been applied into CT image analysis in order to achieve high sensitivity and specific diagnosis results for clinical examinations. In this work, we obtained CT images from a database (CQ500), including epidural hemorrhage, cerebral parenchymal hemorrhage and intraventricular hemorrhage. Then, we introduced a deep-learning algorithm based on U-Net model, which was trained to generate image segmentation, providing a calculated accuracy of prediction yield. The results showed that the average intersection ratio of the final model on the test set could reach the value of 0.96. Briefly, artificial intelligence in this work can efficiently improve the analysis of brain CT images, suggesting an important development direction for future medical imaging auxiliary diagnosis.