Abstract: In order to achieve accurate segmentation of multifidus muscle lesion sites in patients with lumbar disc herniation, a Non-local-based segmentation method was proposed for the multifidus muscle.Based on the U-Net network, the ability of the network to extract multi-cleft muscle features was enhanced by constructing a hybrid pooling convolution instead of the traditional convolution module of the encoder to improve the correlation between global and local features and to fuse the high and low dimensional features of the network. Then, a convolutional module consisting of two cascaded convolutions was deployed in the middle of the network. Finally, after a decoder consisting of a Non-local module and the convolution, the performance of the model was improved by introducing an attention mechanism to focus more on the features of the target and suppressing unnecessary features and noise. According to the experimental results, the proposed model demonstrates a significant improvement in performance compared to the classical U-Net algorithm. Specifically, it increases the Dice coefficient by 9.5%, Jaccard similarity coefficient by 11.3%, and reduces Hausdorff Distance by 74.6%.This method improves the segmentation accuracy of the fatty infiltration site of the multifidus muscle, and provides an effective method for segmentation of multifidus lesion sites in patients with lumbar disc herniation.