Aiming at the problem of limited generalization ability of U-Net network to images of different scales and resolutions in image segmentation tasks. In this paper, a multi-scale feature extraction convolutional block is integrated into U-Net to enable the model to simultaneously consider image information at different scales, thus obtaining a more comprehensive and rich feature representation. In order to alleviate the limitation of the influence of the convolution kernel size in the convolution operation, a self-attention mechanism is embedded on the basis of the multi-scale feature learning image alignment network to cross-fertilize and match the feature representations of different images to form a new UNet backbone network, i.e., the LK-CAUNet model. Segmentation effects of different modal training on image alignment techniques are analyzed using T1, T2, T1ce, Flair single modality and a combination of all four modalities used simultaneously. The segmentation performance of LK-CAUNet model is analyzed on the dataset. The DSC metrics of LK-CAUNet model under the combined training of T1+T2+T1ce+Flair modalities are WT=93.56%, TC=89.42%, and ET=83.22% respectively.
