基于PRAU-Net的新冠肺炎CT图像分割研究  

Research of COVID-19 CT Image Segmentation Based on PRAU-Net

作　　者：曾庆鹏 崔鹏 ZENG Qing-peng;CUI Peng(School of Mathematics and Computer Sciences,Nanchang University,Nanchang 330031,China)

机构地区：[1]南昌大学数学与计算机学院,江西南昌330031

出　　处：《计算机技术与发展》2024年第3期133-139,共7页Computer Technology and Development

基　　金：国家自然科学基金项目(62166026)。

摘　　要：针对新冠肺炎CT影像病灶区域小、形状结构差异大和噪声等问题,提出一种基于编解码结构的PRAU-Net医学图像分割方法。首先,在编码阶段使用一种残差Inception注意力卷积模块(Residual Inception Attention,RIA)提取特征,RIA采用残差结构将更深的并行卷积块和通道注意力机制相结合捕获更丰富的特征;其次,将不同尺度的特征通过跳跃连接进行融合,使解码器中特征有更加丰富的全局信息;最后,在解码器中使用全局注意力模块使网络关注相关特征,有效减少了CT影像中噪声的影响。为了验证该方法的有效性,分别在三个数据集(Segmentation dataset nr.2,CC-CCII和COVID19_1110)上进行验证,实验结果表明,该方法比经典方法分割结果更加准确,相较于U-Net等经典分割方法,Dice系数提升了1.12%~14.84%,敏感度提升了0.7%~24.63%。为了进一步提高分割性能,使用生成对抗网络对Segmentation dataset nr.2数据集进行了扩充,并利用PRAU-Net分割方法和多种经典分割网络进行了验证,结果表明,扩充小样本数据集可以有效地提高分割性能,PRAU-Net方法的Dice系数从0.8364上升到了0.8583。Aiming at the problems of small lesion area,large variation in shape structure and noise in CT images of COVID-19,we proposed a parallel residual attention U-Net medical image segmentation method based on encoder-decoder architecture.Firstly,the model extracted features by residual inception attention(RIA)in the encoder.RIA adopted a residual structure to combine deeper parallel convolution block and channel attention mechanisms to capture richer features.Secondly,features of different scales were fused by skip connection to obtain richer global information.Lastly,the global attention module was used in the decoder to enable the network focus on relevant features,which effectively reduced the influence of noise in CT images.To verify the effectiveness of the proposed method,we have conducted experiments on segmentation dataset nr.2,CC-CCII and COVID19_1110.Experimental results show that proposed method is more accurate than the classical methods.Compared with classical segmentation methods such as U-Net,the Dice coefficient increases by 1.12%~14.84%and the sensitivity increases by 0.7%~24.63%.To further demonstrate the segmentation performance,the Segmentation dataset nr.2 is extended by using generative adversarial network,the PRU-NET method and several classical segmentation networks are used to verify the method.It is showed that expanding the small sample dataset can effectively improve the segmentation performance,the Dice coefficient of PRAU-Net method was increased from 0.8364 to 0.8583.

关 键 词：新冠肺炎 医学图像分割 U-Net 残差结构 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]