基于FPA-U-Net-GAU的肺部病变区域分割方法研究

Research on Segmentation Method of Lung Lesions Based on FPA-U-Net-GAU

针对现有机器学习和深度学习网络算法对肺部病变区域分割精度不高的问题,提出了一种基于改进的特征金字塔注意力机制(Feature Pyramid Attention,FPA)与U-Net网络结合肺部病变区域分割算法。该算法在U-Net采样中嵌入FPA,FPA对目标区域空间位置信息进行分级特征提取,并结合全局池化学习表示特征。在每个上采样模型中引入全局注意力上采样模块(Global Attention Upsample,GAU)作为全局上下文信息的特征提取,网络方法简写为FPA-U-Net-GAU。将FPA-U-Net-GAU与几种经典分割算法模型进行对比,实验结果表明,FPA-U-Net-GAU方法有效提高了分割精度和稳定性,其DSC为0.885,PPV为0.851,Sensitivity为0.882。

Aiming at the problem that the existing machine learning and depth learning network algorithms have low segmentation accuracy for lung lesions,a lung lesions region segmentation algorithm based on improved Feature Pyramid Attention(FPA)and U-Net network is proposed.The algorithm embeds FPA in U-Net sampling,and FPA extracts hierarchical features from the spatial location information of the target area,and combines with global pooling to learn better feature representation.The Global Attention Upsample(GAU)module is introduced into each upsampling model as the feature extraction of global context information.The network method is abbreviated as FPA-U-Net-GAU.Comparing FPA-U-Net-GAU with several classical segmentation algorithm models,the experimental results show that the FPA-U-Net-GAU method effectively improves the segmentation accuracy and stability.Its DSC is 0.885,PPV is 0.851,and sensitivity is 0.882.