【摘要】 针对乳腺钼靶图像中肿块体积小且常被致密组织掩盖导致肿块分割精度较低的问题,提出一种基于复合加权损失函数的U型对称残差语义分割模型SRes-Unet:首先将含有残差结构的卷积模块嵌入U型网络架构中,提升模型整体的特征提取能力;其次,为了解决乳腺图像中因背景较大造成像素类别严重不平衡问题,利用复合型w BCE_DiceLoss作为残差U型网络(SRes-Unet)的损失函数,同时辅以数据增广以减小过拟合风险。实验结果表明,所提分割模型对于乳腺肿块图像能够实现良好的分割效果,DSC值与MIoU值分别达到了0.82和0.86,对比传统U-Net,在DSC和MIoU指标上分别提升了2个百分点和4个百分点。更多还原

【Abstract】 In order to solve the problem of low precision of mass segmentation in breast molybdenum target images,this paper proposed a U-shaped symmetric residual semantic segmentation model named SRes-Unet,based on a compound weighted loss function. First,the residual module was embedded in U-Net to improve the feature extraction ability of the network model. Then,the compound w BCE_DiceLoss was used as the loss function of residual U-type network( SRes-Unet) to solve the problem of serious imbalance of segmented pixel categories caused by large background in breast images,and the data augmentation was used to reduce the risk of overfitting. The results show that the proposed segmentation model can achieve a good segmentation effect for breast masses,with DSC and MIo U reaching 0. 82 and 0. 86,respectively.Compared with U-Net,the DSC and MIo U indicators are improved by 2% and 4%,respectively.更多还原