基于Half-UNet的乳腺肿块分割方法

Breast mass segmentation method based on Half-UNet

乳腺癌是常见的高发病率肿瘤疾病,乳腺肿块分割是乳腺肿瘤分析的重要步骤.为了在保证乳腺肿瘤分割精度的同时提高分割效率,提出了一种基于Half-UNet的乳腺X线摄片图像分割方法.该方法保留了U-Net中分而治之的部分,简化了特征融合的方式.固定U-Net编码器各步骤的特征图通道数以减少网络复杂度的同时有利于解码器的特征融合,并对编码器中的卷积操作增加了he_normal和L2正则化,提升网络性能且缓解网络的过拟合现象.对U-Net解码器的网络结构进行简化,减少网络模型的参数量和训练时间.在DDSM数据集上的实验结果表明:Half-UNet在获得与U-Net、UNet3+模型相近的分割精度的情况下,训练时间相对于U-Net和UNet3+缩短了41.66%和83.33%,显著提升了分割效率.

Breast cancer is a common high incidence rate of cancer.Breast mass segmentation is an important step in breast tumor analysis.In order to improve the segmentation efficiency and ensure the accuracy of segmentation,a Half-UNet based mammography image segmentation method is proposed.In this model,the divide and conquer part of U-Net is preserved,and feature fusion part is simplified.The number of feature map channels in each step of U-Net encoder is fixed to reduce the network complexity,which is beneficial to the feature fusion of decoder,and added he_normal and L2 regularization to the convolution operation in the encoder to improve network performance and alleviate network overfitting.The decoder part of U-Net is simplified to reduce the amount of parameters and computation of the network model.The experimental results on DDSM data sets show that Half-UNet significantly improves the segmentation efficiency.Compared with U-Net and UNet3+,the training time of Half-UNet is reduced by 41.66%and 83.33%,while the segmentation accuracy is almost the same.