To improve the lightweight level and segmentation accuracy of the medical image segmentation model,the Transformer with sparse self-attention calculation method,the boundary segmentation enhancement mechanism and the complementary atten-tion mechanism were introduced to enhance detail feature extraction on the basis of TransUNet,and the original conventional convolution and the upsampling of TransUNet were replaced through deep separable convolution and CARAFE modules.A boundary-accurate segmentation model LB-TransUNet with relative lightweight was designed.Experimental results on the Synapse multi-organ segmentation dataset show that the Dice coefficient and the Hausdorff distance of LB-TransUNet can reach 79.30 and 21.03％respectively,realizing more accurate segmentation effects compared with that of TransUNet,Swin-UNet and other models.
维普
万方数据