A Segmentation Algorithm for Cerebral Hematoma Based on Spatial and Multi-scale Positional Attention Mechanism and Unet

For segmenting the target lesions accurately in CT images of cerebral hematoma, this paper proposes a spatial and multi-scale neural network segmentation model (MSHA-Unet). The model fuses the semantic information of the top and bottom layers of images in the same level by multi-kernel convolution, and fuses the attention mechanism in multi-scale feature images by jumping connection, so as to accurately locate and segment cerebral hematoma in medical CT images. To verify the effectiveness of the model, the experiment was carried out on 3716 CT images of 81 cases of cerebral hemorrhage, and the model was compared with six neural network models. The experimental results were measured by various evaluation standards, and showed that the model was superior to other network models.