| 摘 要: 针对肺结节图像存在体积小、形状不一且不规则、边缘模糊,导致模型特征提取不易及分割精度不高等问题,提出了一种基于TransU-Net结合注意力机制的肺结节分割方法。该方法在 TransU-Net网络架构的基础上引入了改进后的卷积核注意力机制,并提出了一种坐标注意力机制,使特征图获得了更大的感受野,可以更好地提取图像纹理信息,从而提升分割精度。通过在LIDC(Lung Image Database Consortium)肺结节公开数据集上的训练与验证结果表明,所提模型的精确率、召回率、相似系数及均交并比达到了96.90%、95.46%、93.34%和97.30%,证明了该方法的有效性和优越性。 |
| 关键词: 图像分割;注意力机制;肺结节 |
|
中图分类号: TP391
文献标识码: A
|
|
| A Pulmonary Nodule Segmentation Method Based onTransU-Net |
|
YANG Yun1, LIU Jianchen1, YANG Hong2, WU Yanan3
|
(1.School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, X'i an 710021, China;
2.X'i an Friendship Medical Electronics Co., Ltd., X'i an 710075, China;
3.X'i an Institute of Metrology, X'i an 710068, China)
yangyun7021@163.com; 993430612@qq.com; yanghong00012024@163.com; wuyanan0716@163.com
|
| Abstract: To address the challenges in segmenting pulmonary nodules, where small sizes, irregular shapes, and blurred edges lead to feature extraction dificulties and low segmentation accuracy,this paper proposes a pulmonary nodule segmentation method based on TransU-Net integrated with an attention mechanism. The method introduces an improved convolutional kernel attention mechanism into the TransU-Net architecture and proposes a coordinate attention mechanism, enabling the feature maps to capture larger receptive fields and better extract image texture information, thereby enhancing segmentation accuracy. Experimental results on the public LIDC (Lung Image Database Consortium) pulmonary nodule dataset demonstrate that the proposed model achieves precision, recall, similarity coefficient, and mean Intersection over Union (mIoU) scores of 96.90% , 95.46% , 93.34% , and 97.30% respectively.These results validate the effectiveness and superiority of the proposed method. |
| Keywords: image segmentation; attention mechanism; pulmonary nodules |
