摘要： 利用卷积神经网络（CNN）进行医学图像分割时，通常将分割问题抽象为特征表示和参数优化问题，但在上采样和下采样过程中容易丢失特征信息，导致分割效果不理想。设计包含三级特征表示层和特征聚合模块的深度特征聚合网络结构DFA-Net。通过三级特征表示层提取基础特征同时聚合中间特征和深层特征，从而以聚合深层特征弥补CNN上采样与下采样的特征损失。利用特征聚合模块聚合并激活浅层特征和深层特征，根据两者的互补信息分别做精细化调整。在脑图像和眼底图像公开数据集上的实验结果表明，DFA-Net能够充分利用深层特征与浅层特征的信息互补性处理分割结果中的孤立像素点，避免上采样与下采样引起的信息损失，其分割精度较U-net、Unet++、SegNet和LadderNet等方法均有所提高。

关键词: 脑图像分割, 眼底图像分割, 特征聚合, 特征表示, 卷积神经网络

Abstract: Medical image segmentation using Convolutional Neural Network(CNN) usually simplifies the segmentation problem into the feature representation and parameter optimization problems.However,feature information is easily lost in the process of up sampling and down sampling,which leads to unsatisfactory segmentation effect.To solve the problem,this paper designs a deep feature aggregation network structure,DFA-Net,which includes the three-level feature representation layer and feature aggregation module.The three-level feature representation layer is used to extract basic features and aggregate middle features and deep features,so as to use the aggregated deep features to make up for the feature loss in CNN up sampling and down sampling.The feature aggregation module is used to aggregate and activate shallow features and deep features,and perform fine adjustment on them according to their complementary information.Experimental results on open datasets of brain images and fundus images show that DFA-Net can make full use of the information complementarity between deep features and shallow features to deal with isolated pixels in segmentation results and avoid the information loss caused by up sampling and down sampling.Its segmentation accuracy is higher than that of U-net,Unet++,SegNet,LadderNet and other methods.

Key words: brain image segmentation, fundus image segmentation, feature aggregation, feature representation, Convolutional Neural Network(CNN)

中图分类号: 

• TP391