基于残差注意力机制的肺结节数据增强方法  

Data Augmentation of Lung Nodule Based on Residual Attention Mechanism

作　　者：李阳[1,2] 李春璇[2] 徐灿飞 方立梅 LI Yang;LI Chunxuan;XU Canfei;FANG Limei(Academy for Advanced Interdisciplinary Studies,Northeast Normal University,Changchun 130000;School of Computer Science and Engineering,Changchun University of Technology,Changchun 130012;Run Run Shaw Hospital,Zhejiang University School of Medicine,Hangzhou 310000;The Third Affiliated Clinical Hospital of Changchun,University of Traditional Chinese Medicine,Changchun 130117)

机构地区：[1]东北师范大学前沿交叉研究院,长春130000 [2]长春工业大学计算机科学与工程学院,长春130012 [3]浙江大学医学院附属邵逸夫医院,杭州310000 [4]长春中医药大学附属第三临床医院,长春130117

出　　处：《电子科技大学学报》2023年第6期880-886,共7页Journal of University of Electronic Science and Technology of China

基　　金：国家自然科学基金(NSFC 61806024,NSFC12226003,NSFC12171076);吉林省科技厅重点研发计划(20210201081GX,20200401103GX);吉林省教育厅重点科研项目(JJKH20220685KJ,JJKH20220692KJ);吉林省卫生厅卫生科研人才专项(2020SCZ25)。

摘　　要：针对带标注的肺CT图像数据匮乏而导致的深度学习模型训练困难,以及现有生成算法生成肺结节不同特征模糊、细节丢失的问题,提出了肺结节图像的数据增强RAU-GAN算法。首先,在生成器网络中嵌入残差注意力模块,该模块可以聚焦于局部不同的感兴趣区域,以实现肺结节与背景信息的独立生成,并且重新设计了注意力模块中的残差块来减少网络的深度和训练的复杂度。其次,将判别器设计为U-Net架构,可以给更新后的生成器反馈更多信息,以提高判别性能。最后,在数据集LUNA16和Deep Lesion上进行实验,结果与现有方法相比,在视觉效果和不同评价指标上均有提升,验证了生成图像包含了更丰富的细节信息。Aiming at the difficulty of deep learning model training caused by the lack of labeled lung Computed Tomography(CT)image data and the lung nodule feature model generated by existing generation algorithmsTo solve the problem of blur and detail loss,a data-enhanced RAU-GAN algorithm for pulmonary nodule images is proposed.Firstly,a residual attention module is embedded in the generator network,which can focus on different local regions of interest to achieve the independent generation of lung nodules and background information.Moreover,the residual block structure in the attention module is redesigned to to reduce the depth of the network and training complexity.Second,the discriminator is designed as U-Net architecture,which can feed back more information to the updated generator to improve the discrimination performance.Finally,experiments were conducted on data set LUNA16 and deep lesion.The results show that the visual and different evaluation indexes have improved in comparison with existing methods,which verifies that the generated images can contain richer details.images can contain richer details.

关 键 词：数据增强 Pix2Pix RAU-GAN 残差注意力机制 U-Net判别器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]