%0 Journal Article
%T 关联自注意力机制改进的UNet用于肺结节图像分割<br>Associative Self-Attention Mechanism Improved UNet for Lung Nodule Image Segmentation
%A 周稳
%A 王继红
%J Modeling and Simulation
%P 5522-5530
%@ 2324-870X
%D 2023
%I Hans Publishing
%R 10.12677/MOS.2023.126501
%X 肺癌在各项肺部疾病中是非常致命的，而早期的诊断和治疗则能够有效的降低死亡率，对于早期的预防以及诊断，肺结节分割是一个非常重要的步骤，但结节与其领域之间的相似性难以处理。在这里本文提出了一个端到端的基于UNet的分割框架，称为PR-Net。为了保证标准卷积的可接受精度以及深层结构带来的信息损失，本文构建了残差结构的变体作为网络的主干。此外在解码过程中由于信息的冗余和语义差距导致难以有效的融合病变区域的空间及语义信息，为此本文采用了Transformer中的注意力机制，通过该机制的多头自注意力来有效提升空间细节和语义定位级别的特征识别。最终的模型IOU指标达到了97.71%，Dice相似系数达到了98.84%。大量的实验和结果表明，本文的模型在肺结节分割上有着非常不错的性能及稳定性。<br />
Lung cancer is very fatal among all lung diseases, while early diagnosis and treatment can effective-ly reduce the mortality rate. For early prevention as well as diagnosis, lung nodule segmentation is a very important step, but the similarity between the nodule and its domain is difficult to handle. Here in this paper we propose an end-to-end UNet-based segmentation framework called PR-Net. To ensure acceptable accuracy of standard convolution as well as loss of information due to deep structure, a variant of residual structure is constructed as the backbone of the network. In addition, in the decoding process, the redundancy of information and semantic gaps lead to the difficulty of effectively fusing the spatial and semantic information of the lesion region, for this reason, this pa-per adopts the attention mechanism in Transformer, and effectively improves the feature recogni-tion at the spatial detail and semantic localization level through the multi-head self-attention of this mechanism. The final model IOU index reaches 97.71% and Dice similarity coefficient reaches 98.84%. A large number of experiments and results show that the model in this paper has very good performance and stability in lung nodule segmentation.
%K 深度学习，图像分割，UNet，自注意力<br>Deep Learning
%K Image Segmentation
%K UNet
%K Self Attention
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=75865