基于空洞U-Net网络的乳腺细胞图像分割算法

Breast cell image segmentation algorithm based on dilated U-Net network

乳腺细胞的准确分割是乳腺组织切片图像病理分析的关键环节,对乳腺癌的诊治具有重要价值。针对乳腺细胞图像分割中细胞边界不清晰、分割准确率低的问题,提出一种基于空洞U-Net网络的乳腺细胞图像分割算法。在U-Net网络中引入空洞卷积增大卷积层感受野,获得包含更多乳腺细胞边缘信息的特征图,在卷积层和池化层间增加实例归一化层,提高网络收敛速度的同时缓解过拟合现象,并使用加权损失函数增强乳腺细胞区域的权重,提高网络对细胞特征的提取能力,实现乳腺细胞边界的有效分割。实验结果表明,该算法在USCB Breast数据集上的分割准确率和Dice系数分别达到97.63%和83.25%,较原始U-Net网络分别提高了6.5%和6.6%,对乳腺细胞图像具有更好的分割效果。

Accurate segmentation of breast cells is the key link in the process of pathological analysis of breast tissue section images,which has important value for the diagnosis and treatment of breast cancer.To solve the problems of unclear cell boundary and low segmentation accuracy in breast cell image segmentation,an algorithm of breast cell image segmentation based on dilated U-Net network is proposed.Dilated convolution is introduced into U-Net network to increase the receptive field of convolution layer to obtain the feature maps containing more breast cell boundary information.An instance normalization layer is added between the convolution layer and the pooling layer to improve the convergence speed of the network while alleviating the over fitting phenomenon.The weighted loss function is used to enhance the weight of breast cell region and improve the ability of the network to extract cell features,so as to realize the effective segmentation of breast cell boundaries.The experimental results on the USCB Breast dataset show that the segmentation accuracy and Dice coefficient of the proposed algorithm are 97.63%and 83.25%,respectively,which are 6.5%and 6.6%higher than that of the original U-Net network,and the segmentation effect is better for breast cell images.