After prostate magnetic resonance (MR) image slices,it is found that some images do not have effective edge information,which makes it impossible to clea rly locate the edge position,and thus cannot segment the prostate.At the same time ,the traditional convolutional network requires a large amount of parameters and take s up too much storage space of the model.This paper proposes a method to segment the pro state using U-Net that combines multi-scale dilated separable convolution and channe l attention.First,slice 50three-dimensional (3D) prostate samples and perform contrast enhancement on the sliced images.Subsequently,the processed data is input into the residual U-Ne t,and the multi-scale dilated convolution and channel attention are used as the encoding -decoding unit to extract the feature information.Finally,the Dice coefficient and Hausd orff distance (HD) are used to evaluate the segmentation results.The experiment was verified on the PROMISE12challenge dataset,and the final Dice coefficient and HD were 88.13% and 14.17mm,respectively,and the parameter amount and storage space wer e reduced by 57%.The results show that this method can not only segment the pros tate area without effective edges to improve its segmentation accuracy,but also effe ctively reduce the parameter amount and storage space,and can be applied to medical ima ges with blurred edges.