作者:刘宇博，刘国柱，史操，许灿辉
\n

Authors:LIU Yu-bo,LIU Guo-zhu,SHI Cao,XU Can-hui
摘要:摘要:针对传统的计算机辅助诊断系统对低剂量肺部CT图像结节检出率低、假阳性高等问题，提出一种类U-Net网络和基于注意力机制的两阶段肺结节检测模型。为了提高肺结节的检测速度和检出率，首先构建了一个三维网络用于候选结节的检测，充分利用结节的三维信息提高候选结节的检出率的同时，优化了检出速度；然后采用多视图输入方式以保证对结节空间特征的获取，将结节在三维空间下的矢状面、冠状面、水平面等9个角度下的切片一起输入网络，利用ViT模型做特征提取器并结合特征金字塔网络实现对结节的分类，将所有切片结果融合以实现对假阳性结节的筛除。最终在LUNA16数据集上的实验结果表明，所提出的模型准确率达到94.7%，提高了准确率的同时降低了误诊率和漏诊率。
Abstract:Abstract:To solve the problems of low detection rate and high false positive of nodules in low-dose lung CT images by traditional computer-aided diagnosis system, a two-stage pulmonary nodules detection model based on U-Net network and attention mechanism was proposed. In order to improve the detection speed and detection rate of pulmonary nodules, a 3D network was constructed to detect the candidate nodules firstly. It optimized the detection speed while the three-dimensional information of nodules was fully utilized to improve the detection rate of the candidate nodules. Then, the multi-view input method was used to ensure that the spatial features of nodules was obtained. The sections from 9 angles in three-dimensional space, including sagittal plane, coronal plane and horizontal plane, were input into the network together.The ViT network was used as a feature extractor and combined with the feature pyramid network to achieve the classification of nodules, and we fused all section results to achieve the screening of false positive nodules. The final experimental results on LUNA16 data set show that the accuracy of the proposed model reaches 94.7%, which improves the accuracy and reduces the rate of misdiagnosis and missed diagnosis.

PDF全文下载地址:

可免费Download/下载PDF全文