摘要
CT影像学检查是诊断肺癌的主要手段,对CT图像的肺部图像进行分割可以辅助医生诊断肺部疾病。针对现有算法在肺部分割上精度不够,细节处理不完善的问题,提出了一种全新的编码-解码结构的TCU-Net分割方法,使用混合变压器结构作为编码来扩大感受野,利用多头自注意力机制来捕获全局的上下文信息从而对目标建立起远距离的依赖,提取出更强有力的特征;使用转置卷积恢复其空间分辨率,减少特征的损失;在编码和解码之间添加CE模块,该模块可以捕获更多抽象特征和空间信息。实验结果表明,相比U-Net算法在精确率、Dice系数、敏感度、F_(1)分数上分别提高了0.48%、0.8%、0.28%、1.49%。
CT imaging examination is the main means to diagnose lung cancer.The segmentation of lung CT images can help doctors diagnose lung diseases.Aiming at the problems of insufficient precision and imperfect detail processing of existing algorithms in lung segmentation,a new encode-decode lung segmentation method of TCU-Net was proposed.A hybrid Transformer model was used as the code to enlarge the receptive field,multi-attention mechanism is used to capture global context information and establish remote dependence on the target to extract more powerful features.Transpose convolution is used to restore its spatial resolution and reduce the loss of features.CE Block is added between encoding and decoding,which can capture more abstract features and spatial information.Experimental results show that compared with U-Net algorithm,Accuracy,Dice similarity coefficient,Sensitivity and F_(1)-score are improved by 0.48%,0.8%,0.28%and 1.49%respectively.
作者
穆应晨
王学军
王怡雯
卢琳
MU Yingchen;WANG Xuejun;WANG Yiwen;LU lin(College of Information Sciences and Technology,Shijiazhuang Tiedao University,Shijiazhuang 050043,China)
出处
《安徽工程大学学报》
CAS
2022年第1期39-45,共7页
Journal of Anhui Polytechnic University
基金
石家庄铁道大学研究生科研创新基金资助项目(YC2021076)。
关键词
肺部
编码器
混合变压器
图像分割
lung
encoder
hybrid transformer
image segmentation
