摘要
为了提高对肺结节的准确分割,文中提出一种CSF-UNet的双骨干网络特征提取方法。使用两种不同侧重的骨干网络并行提取图像特征,通过利用ConvNeXt网络提取局部特征,并结合Swin Transformer网络提取全局特征来提升模型的特征提取能力。提出了一种自适应大核融合模块,有效地融合两种不同规格的特征,通过串联两个大核卷积获得更大的感受野和动态选择机制来突出重要的空间区域。在SPPF中融合了ECA通道注意力和密集链接,提出了ESPP模块以进一步挖掘双骨干网络提取的高级语义信息,使网络更加关注重要的特征通道。在LIDC数据集上的实验结果表明,提出的模型在3个指标上优于基本UNet模型以及最近几年其他研究团队提出的针对该数据集的分割网络。最终,CSF-UNet模型实现了78.1%的IoU、87.71%的DSC、87.19%的敏感度和88.23%的精确度。这些结果表明,该模型在肺结节分割方面表现出良好的性能,对医生进行早期肺结节诊断具有重要的临床意义和应用价值。
In order to enhance the segmentation accuracy of pulmonary nodules,this study proposes a dual-backbone network feature extraction method named CSF-UNet.Two backbone networks with different emphases are used to extract image features in parallel.ConvNeXt is employed to capture local features,while Swin Transformer is utilized to extract global features,so as to enhance the feature extraction capabilities of the model.An adaptive large kernel fusion module is introduced to integrate features of different scales effectively.By concatenating two large kernel convolutions,a larger receptive field and a dynamic selection mechanism are achieved to highlight important spatial regions.The ECA(efficient channel attention)and dense connections are integrated into SPPF(spatial pyramid pooling fusion),and an ESPP module is proposed to further exploit the high-level semantic information extracted by the dual-backbone networks,so as to make the network focus on critical feature channels.Experimental results on the LIDC dataset demonstrate that in terms of the three indicators the proposed model outperforms the baseline model UNet and other recent segmentation networks developed for this dataset and proposed by other research teams.Ultimately,the CSF-UNet model achieves IoU(intersection over union)of 78.1%,DSC(dice similarity coefficient)of 87.71%,sensitivity of 87.19%and precision of 88.23%.These results indicate that the proposed model exhibits robust performance in pulmonary nodule segmentation,holding significant clinical implications and application value for the diagnosis of early-stage pulmonary nodule.
作者
郝胜男
庞建华
HAO Shengnan;PANG Jianhua(School of Artificial Intelligence,North China University of Science and Technology,Tangshan 063210,China)
出处
《现代电子技术》
北大核心
2025年第1期1-7,共7页
Modern Electronics Technique
基金
国家重点研发计划(2017YFE0135700)基金。
