For the phenomenon that the kidney tumor target is small and difficult to segment, and the sample distribution is uneven, 3*3 residual convolution blocks are used to replace the 3*3 convolution of the original U-Net model. The problem of gradient disappearance is avoided while stabilizing the number of deepening layers to improve the effect of the model. The NAM (Normalization Attention Module) module is embedded in the upsampling process to enhance the network's ability to extract features. The proposed network achieves 0.917, 0.938, 0.921, 0.960 and 0.887 in terms of recall, Dice coefficient, IOU, F1 score and accuracy compared with other networks. It can be seen that the method proposed in this paper has improved the situation of small target misdetection and omission to a certain extent, and can effectively segment the kidney and tumour.