融合Canny边缘检测的多输出损失肺炎CT图像分割算法

Multi⁃output loss pneumonia CT image segmentation algorithm fused with Canny edge detection

针对肺炎的影像学特征如弥漫、多灶、磨玻璃的特点,提出一种肺部感染CT图像分割算法CEDMO。算法主干网络使用ResNet50,在特征提取中,先利用Canny算子对分割目标边缘进行检测,得到的边缘信息再与ResNet50进行融合计算,边缘信息还在解码器部分的多目标输出计算中作为一个引导值。在解码器部分对PSA注意力机制改进并设计了多输出约束,以获得了更多的细节信息和多个尺度的特征。在多输出中,设计了5个输出路径,每个路径的Loss都参与约束计算,使得训练结果加快收敛,从而提高计算效率。最后通过实验对比基线模型的结果,3个指标优于基线模型,其中Dice系数、灵敏度(SE)、增强对准度量(E_?~m)依次提高了4.7、4.5和1.3个百分点。

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(E^(m)_(ϕ))increased by 4.7,4.5 and 1.3 percentage,respectively.