基于KCCS-YOLOv4的肺结节检测算法  

Pulmonary nodule detection algorithm based on KCCS-YOLOv4

作　　者：陈茂龙 李亚洲 云家伟 程照雪 李阳[1] CHEN Maolong;LI Yazhou;YUN Jiawei;CHENG Zhaoxue;LI Yang(School of Computer Science&Engineering,Changchun University of Technology,Changchun 130102,China)

机构地区：[1]长春工业大学计算机科学与工程学院,吉林长春130102

出　　处：《长春工业大学学报》2023年第5期424-433,共10页Journal of Changchun University of Technology

基　　金：吉林省教育厅重点科研项目(JJKH20220685KJ)。

摘　　要：针对单阶段目标检测算法在肺结节检测任务中存在易错检及漏检、检测精度偏低、计算复杂度高等问题,提出了一种以YOLOv4为主干框架,并结合K-means++聚类算法、坐标注意力机制、卷积块注意力模块和SIoU Loss的KCCS-YOLOv4肺结节检测算法。采用K-means++聚类算法针对肺结节进行锚框优化,以降低漏检概率及计算复杂度。在YOLOv4的特征融合模块中引入坐标注意力机制以辅助检测网络捕获结节的位置信息和空间信息,引入卷积块注意力模块以辅助检测网络捕获在上、下采样过程中丢失的特征信息。采用SIoU Loss优化损失函数,提高训练的收敛速度和肺结节的检测精度。在公开数据集LUNA16上进行测试和分析,实验结果与四种现有算法进行比较,所提KCCS-YOLOv4算法的F 1-score为93.7%,平均精度值为95.9%,FLOPs值为127.592 G,三项指标均优于其他算法,验证了所提改进策略的有效性。Aiming at the problems of single-stage target detection algorithm in the detection task of pulmonary nodules,such as error-prone and missed detection,low detection accuracy,and high computational complexity,a framework based on YOLOv4 was proposed,combined with K-means++clustering algorithm,Coordinate attention mechanism,Convolution block attention module and SIoU Loss's KCCS-YOLOv4 lung nodule detection algorithm.First,the K-means++clustering algorithm is used to optimize the anchor frame for pulmonary nodules to reduce the probability of missed detection and computational complexity.Secondly,the coordinate attention mechanism is introduced in the feature fusion module of YOLOv4 to assist the detection network to capture the position information and spatial information of nodules,and the convolution block attention module is introduced to assist the detection network to capture the lost features in the process of upsampling and downsampling information.Then,SIoU Loss is used to optimize the loss function to improve the convergence speed of training and the detection accuracy of pulmonary nodules.Finally,the test and analysis were carried out on the public data set LUNA16,and the experimental results were compared with four existing algorithms.The F 1-score of the proposed KCCS-YOLOv4 algorithm was 93.7%,the average precision was 95.9%,and the FLOPs value was 127.592 G,the three indicators are better than other algorithms,which verifies the effectiveness of the proposed improvement strategy.

关 键 词：肺结节检测 YOLOv4 K-means++ 注意力机制 SIoU Loss 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]