面向病理图像分割的边缘感知网络

Boundary Perception Network for Pathological Image Segmentation

提出了一种针对病理切片图像的端到端语义分割方法--边缘感知网络(BPNet),以提高病理图像分割精度。BPNet网络首先在解码器阶段增加边缘感知模块,改善网络对于病理图像边缘的特征信息提取能力。然后,采用自适应通道注意力模块弥补不同层次特征间的语义差距,进一步加强网络的特征聚合能力。在此基础上,设计了一种基于结构和边缘的联合损失函数,以实现最佳的病理图像分割结果。在GlaS和MoNuSeg两个公开病理数据集上的分割实验结果表明,所提方法的Dice系数得分在两个数据集上分别达到92.21%和81.18%,有效提升了病理图像的分割精度。

In clinical practice,segmentation and quantitative evaluation of target objects in pathological images provide valuable information for histopathological analysis,which is of great significance to auxiliary diagnosis and subsequent treatment.However,due to the dense distribution of cells and great morphological similarities between the cancer cells and normal cells,there are some challenges such as difficulty in feature extraction and unclear segmentation boundaries in the segmentation task of pathological images.At the same time,the traditional image segmentation methods are time-consuming and laborintensive.They can only extract low level manual features,and the expression ability of deep discrimination features is insufficient,resulting in limited performance of traditional methods.Meanwhile,previous deep learning algorithms still suffer from two significant problems.Firstly,most networks ignore pixels that are difficult to segment,such as the boundaries of targets,which is particularly important for accurate segmentation.In addition,the problem of inconsistent semantic levels between different features are not solved,leading to low training efficiency.To address the above-mentioned problems,an end-toend histopathological image segmentation network called Boundary Perception Network(BPNet)is proposed for improving the segmentation accuracy of histopathological images.Based on encoder-decoder structure,the encoder performs the convolutional downsampling operation to extract the feature information of the image through the Convolutional Neural Network(CNN).And the encoding process uses the feature encoder based on the EfficientNet-B4 network which is specifically used for pathological image segmentation.The decoder mainly consists of decooding blocks,Boundary Perception Module(BPM)and Adative Shuffle Channel Attention Moudule(ASCAM).In detail,the decoding block performs deconvolution operation to complete the decoding process of the feature information.Then,the BPM in the decoder stage aims to strengthen the ability of mining for difficult segmentation regions,so that the network focuses on the higher uncertainty as well as more complex edge regions,achieving feature complementarity and precision prediction results.For implementation,the BPM extracts the edge from the decoder output of each layer,and superimposes the edge information onto the encoded feature to strengthen the boundary feature information extracted from pathological images,outputting the enhanced edge perception feature map.Subsequently,the ASCAM is an improved chanel attention moudule which is used to make up the semantic gap between different levels of features,extrated by encoder,decoder and BPM,so as to further strengthens the feature understanding ability of the BPNet.This module exploits adaptive kernel size one-dimensional convolusion to capture the interactive information of local channels,at the same time ensures the efficiency and effectiveness of the training process.The obtained channel attention coefficient is multiplied by the module input feature layer to obtain the fusion feature,helping effectively learn the channel interaction information between features to improve the feature representation ability.Furthermore,a joint loss function based on structure and boundary is designed to optimize the targeting and detail processing capabilities of this method,achieving the better segmentation result of pathological images.Experiments are carried out on the Gland segmentation(GlaS)and MoNuSeg dataset,respectively.Both of the two datasets are devided into 4∶1 for training and validation.At the same time,in order to make up for the overfitting caused by the lack of training data,two kinds of online data enhancement methods of horizontal flipping and vertical flipping were carried out on the training set data in the experiment.And the four evaluation index,the Dice coefficient score,Intersection Over Union(IoU),Accuracy(ACC)and Precision(PRE),are used to evaluate the performance of this method propsed in this paper.The Dice coefficient score of the proposed method is 92.21%and 81.18%,the IoU is 85.55%and 68.34%,the ACC is 92.14%and 92.50%,the PRE is 92.07%and 75.46%on the GlaS and MoNuSeg datasets,respectively.Compared with the previous classical methods,such as U-Net,UNet++,MultiResUNet,TransUNet,UCTransNet and so on,the BPNet proposed gets the best segmentation result,especially retains more details in the segmentation boundary.Moreover,ablation experiments are carried out on the same two datasets for indicating the impacts of BPM and ASCAM.The results shows that the proposed BPM significantly optimizes the segmentation effect of the network for the edge,as well as the ASCAM makes up the semantic gap between features at different levels and further strengthens the feature understanding ability of the network.In conclusion,the BPNet proposed in this paper exploits BPM to generate edge enhancement feature maps,and uses ASCAM to seize crucial features.Finally,a joint loss function is used to capture the information of features at different levels in the output layer to achieve optimal segmentation performance.The experimental results have demonstrated that the effectiveness of each part of proposed method in the segmentation task of pathological images.