Brain hemorrhage is a threatening disease with a yearly increasing incidence. Computed tomography (CT) is a common method of obtaining hematoma information and periodically monitoring changes of brain injuries. However, due to high frequency tomography, a large number of CT images is acquired, which complicates the analysis process. To increase the speed of analysis and ensure the accuracy of CT detection, we combined CT with deep learning to obtain automatic segmentation. In the present study, we developed a segmentation model based on a U-net with residual effects, for hemorrhage images. First, we screened the data and separated it into three parts for training, evaluation, and blind testing. Second, we pre-processed the dataset for data augmentation, which was used to avoid overfitting. After data augmentation, we transferred the data to an algorithm for training. As for the final model, we obtained an image segmenter with a mean intersection over union score of 0.8871 and dice score of 0.9362. The velocity of this algorithm was 26.31 fps, which greatly increased the speed of analysis. Thus, the segmenter obtained high detection efficiency and quantitative detection, which was suitable for periodically monitoring the areas of bleeding and assisting physicians in developing therapeutic regimens. Furthermore, the binary segmentation algorithm can be used for the development of pretraining models for classified segmentation tasks of CT images of head hemorrhages.