基于数据增强及注意力机制的肺结节检测系统

带标注的医学影像数据过少使模型的学习能力有限，深度检测过程中下采样带来的微小结节特征信息容易丢失。为此，设计了一种基于计算机断层扫描的生成式对抗网络（CT-GAN）的数据增强及改进YOLO-V4检测框架的肺计算机辅助检测系统。首先，在结节生成框架CT-GAN中引入DropBlock正则化方法，实现带标注医学影像的数据增强，以提升肺结节的生成质量； 其次，在YOLO-V4中引入坐标注意力机制，以捕捉肺结节的位置感知、方向感知和跨通道的信息，更加精确地检测肺结节感兴趣区域。实验结果表明，在LUNA16数据集上，所提框架的数据增强和结节检测的性能优于其他框架。

Lung Nodule Detection System Based on Data Augmentation and Attention Mechanism

To solve the problem of limited model learning ability caused by insufficient labeled medical image data and easy loss of tiny nodule features caused by sub-sampling in the process of deep detection, a lung computer aided-detection system based on a generative adversarial network based on computed tomography data augmentation and improved you only look once-V4 (YOLO-V4) detection framework is designed. First, the regularization method DropBlock is introduced into the nodule generation framework computed tomography-generative adversarial networks to augment the data of annotated medical images, which can improve the generation quality of pulmonary nodules. Second, the coordinate attention model is introduced in YOLO-V4, which was constructed to capture the position perception, direction perception and cross-channel information of pulmonary nodules, which can further help the model to detect the region of interest of pulmonary nodules more accurately. The experimental results show that the performance indexes of data augmentation and nodule detection in the lung nodule analysis 16 data set of the proposed lung computer aided detection system are superior to the comparison algorithm, which can effectively expand the data set and improve the performance of nodule detection.