In view of the imaging features of pneumonia,such as diffuse,multifocal and ground-glass,a CT image segmentation network for lung infection CEDMO is proposed.The backbone network uses ResNet50,in feature extraction,the Canny operator is first used to detect the edge of the segmented target,and the resulting edge information is then fused with ResNet50 for calculation,and the edge information is also used as a guide value in the multi-target output calculation of the decoder part.In the decoder part,the PSA attention mechanism is improved and the multi-output constraint is designed to obtain more detailed information and fea-tures at multiple scales.In the multiple outputs,five output paths are designed,and the loss of each path participates in the con-straint calculation,so that the training results converge faster,thereby improving the computational efficiency.Finally,by comparing the results of the baseline model experimentally,the three indicators were better than the baseline model,among which the Dice coef-ficient,the sensitivity(SE),and the enhanced alignment metric(Emϕ)increased by 4.7,4.5 and 1.3 percentage,respectively.
维普
万方数据