非对称卷积核YOLO V2网络的CT影像肺结节检测

肺癌一直是严重威胁人类健康的疾病之一,肺结节作为早期肺癌的一个重要征象,在肺癌的早期诊断与治疗中具有重要的意义。传统的CT影像肺结节检测方法不仅步骤繁琐、处理速度慢,而且对于结节的检出率及定位精度都亟待提高。提出一种基于非对称卷积核YOLO V2网络的CT影像肺结节检测方法:首先将连续的CT序列叠加构造为伪彩色数据集,以增强病变和健康组织的差异;然后将含有非对称卷积核的inception V3模块引入到YOLO V2网络中,构造出一种适用于肺结节检测的深度网络,一方面利用YOLO V2网络在目标检测上的优势,另一方面通过inception V3模块在网络的宽度与深度上进行扩增,以提取更加丰富的特征;为进一步提高结节的定位精度,对损失函数的设计与计算方法也进行一定的改进。为验证所提检测模型的性能,从LIDC-IDRI数据集中选取1 010个病例的CT图像用于训练和测试,在大于3 mm的肺结节中,检测敏感度为94.25%,假阳性率为8.50%。实验表明,所提出的肺结节检测方法不仅可以简化肺部CT图像的处理过程,而且在结节检测率及定位精度方面均优于传统方法,可为肺结节检测提供一种新思路。

YOLO V2 Network with Asymmetric ConvolutionKernel for Lung Nodule Detection of CT Image

Lung cancer has always been one of the serious threats to human health. As an important sign of early lung cancer, pulmonary nodules are of great significance in the early diagnosis and treatment of lung cancer. The traditional CT image lung nodule detection method is not only cumbersome, but also slow in processing speed, and the detection rate and positioning accuracy of the nodules need to be improved. This paper proposed a CT image lung nodule detection method based on the asymmetric convolution kernel YOLO V2 network. First, the continuous CT sequence was superimposed to construct a pseudo-color data set to enhance the difference between lesion and healthy tissue, which contained asymmetric volume. The inception V3 module of the accumulation was introduced into the YOLO V2 network to construct a deep network suitable for lung nodule detection. This aspect was drawn on the advantages of the YOLO V2 network in target detection, and on the other hand through the inception V3 module. The width and depth of the network were amplified to extract more abundant features; in order to further improve the positioning accuracy of the nodules, the design and calculation method of the loss function had also been improved. In order to verify the performance of the proposed test model, CT images of 1010 cases were selected from the LIDC-IDRI data set for training and testing. In lung nodules larger than 3 mm, the detection sensitivity was 94.25%, and the false positive rate was 8.50%. Experimental results showed that the lung nodule detection method proposed in this paper not only simplified the processing of lung CT images, but also was superior to traditional methods in nodule detection rate and positioning accuracy, providing a new way for lung nodule detection.