新型架构下的密集网络在肺部影像的分割研究  

Segmentation of Lung Images with Dense Network Models Under a New Architecture

作　　者：陈亮 高文根[1,2] 张晨 陈东[1,2] CHEN Liang;GAO Wengen;ZHANG Chen;CHEN Dong(School of Electrical Engineering,Anhui Polytechnic University,Anhui Wuhu 241000,China;Key Laboratory of Advanced Sensing and Intelligent Control for High-end Equipment,Ministry of Education,Anhui Polytechnic University,Anhui Wuhu 241000,China)

机构地区：[1]安徽工程大学电气工程学院,安徽芜湖241000 [2]安徽工程大学高端装备先进感知与智能控制教育部重点实验室,安徽芜湖241000

出　　处：《重庆工商大学学报（自然科学版）》2023年第6期53-60,共8页Journal of Chongqing Technology and Business University:Natural Science Edition

基　　金：国家自然科学基金项目(61976005).

摘　　要：在医学图像分割领域中,肺实质的分割对肺结节检测有着至关重要的作用,在考虑到模型参数量的情况下追求更高的精度一直是研究热点之一;为此提出了新的三层密集卷积神经网络DA-UNet,首先用密集卷积模块代替在传统U-Net使用的普通3×3卷积,利用密集卷积特征重用特点,加强了网络的特征提取能力。再者在没有太过影响分割网络精确度的前提下加以修剪,减少了上下采样次数,减少不必要的算力消耗。此外,使用了注意力门(Attention gate),加强了跳跃连接中高底层信息融合效果,并且使用空洞空间金字塔池化(Atrous spatial pyramid pooling),模型加入了不同尺度的特征信息,进一步加强图像中任务相关的区域特征,有效减小噪声干扰,提高网络分割精度。通过实验证明:三次上下采样改进模型的参数量只有传统四次上下采样的75.2%左右,但是分割效果没有太大的影响,用LUNA竞赛肺部影像数据集进行了分割验证,实验结果在测试集上的准确率达到了0.991,而IoU则为0.961,比起传统U-Net的评价指标IoU提升了2.9%;在泛化实验的肝脏图像中,DA-UNet的IoU稳定在0.929左右,而U-Net稳定在0.838左右。这些结果证明了改进的U-Net有更佳的分割效果。In the field of medical image segmentation,the segmentation of lung parenchyma plays a crucial role in lung nodule detection.Pursuing higher accuracy while considering the number of model parameters has been one of the research hotspots.In this regard,a new three-layer dense convolutional neural network,DA-UNet,was proposed.Firstly,a dense convolution module was used instead of the normal 3×3 convolution used in traditional U-Net,which enhances the feature extraction capability of the network by taking advantage of the reuse feature of dense convolution.Furthermore,the network was pruned without affecting the accuracy of the segmentation network too much,reducing the number of up and down samples,and reducing unnecessary computational power consumption.In addition,the Attention Gate was used to enhance the effect of information fusion between high and low levels in the jumping connection,and atrous spatial pyramid pooling was used.The model incorporated feature information at different scales to further enhance task-related regional features in the images,effectively reducing noise interference and improving network segmentation accuracy.It was demonstrated experimentally that the number of parameters of the three-time up-and-down sampling improved model was only about 75.2%of those of the traditional four-time up-and-down sampling,but the segmentation effect was not much affected.The LUNA competition lung image data set was used for segmentation verification.The experimental results showed that the accuracy reached 0.991,and the IoU was 0.961.Compared with the traditional U-Net,the evaluation index IoU was improved by 2.9%.In the liver images of the generalization experiment,the IoU of DA-UNet was stable at about 0.929,while that of U-Net was stable at about 0.838.These results prove that the improved U-Net has better segmentation results.

关 键 词：U-Net 密集网络 肺实质分割 空洞空间金字塔池化 注意力门 DA-UNet 评价指标 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]