Abstract: A network model combining self-attentive mechanism， superposition loss function，and U-Net network was proposed to achieve accurate segmentation of neuronal images for the current defects of feature ambiguity，high complexity，and lossy edges of electron microscopy neuronal image segmentation. Firstly，the geometric transformation was performed with the original image to enlarge the data set and reduce overfitting. Secondly，the improved self-attention mechanism was used to focus on learning image details to improve the accuracy of model segmentation. Finally，the network performance was improved by appropriately combining Dice Loss with relative entropy （KL scatter）. The network experimented on the ISBI 2012 dataset， and its correctness， F1 index，accuracy，and recall reached 0. 930 43，0. 956 79，0. 953 26，and 0. 960 34，respectively. The image segmentation effect was relatively more accurate in overall and detail segmentation，and the cell membrane segmentation was basically unbroken.

Key words: U-Net network, neuron segmentation, attention mechanism, KL scatter, BN layer