文献类型：期刊文献

中文题名：改进YOLOv5的肺结节检测算法

英文题名：Lung nodule detection algorithm based on improved YOLOv5

作者：田吉[1];杨萍[1];刘佳[1];王金华[1]

第一作者：田吉

机构：[1]北京联合大学智慧城市学院,北京100101

第一机构：北京联合大学智慧城市学院

年份：2025

卷号：42

期号：1

起止页码：43-51

中文期刊名：中国医学物理学杂志

外文期刊名：Chinese Journal of Medical Physics

基金：国家自然科学基金(62172045,62272049)。

语种：中文

中文关键词：YOLOv5;肺结节检测;下采样算法;注意力机制;困难样本

外文关键词：YOLOv5;lung nodule detection;downsampling algorithm;attention mechanism;hard sample

摘要：针对肺部CT图像中大量小结节难以检测、以及现有肺结节检测算法难以实现轻量化和高精度兼顾的问题,提出改进YOLOv5的高精度轻量化肺结节检测算法。主要改进以下4个方面:(1)使用空间-深度下采样操作替换YOLOv5主干网络中步长为2的下采样操作,使细微特征提取更完整以便于发现微小结节;(2)在YOLOv5颈部使用渐进融合特征策略,构建不同路径特征图之间的联系以增强各个层级之间信息的交互;(3)创造性地提出了感知全局上下文注意力并将其应用在YOLOv5颈部网络的末端,提高模型从全局视角对肺结节关键特征和语义信息的理解能力;(4)采用损失排序挖掘方法重点训练困难样本,以此来强化模型的鉴别能力。改进后的算法在LUNA16数据集上得到了96.0%的精确度,95.0%的召回率和97.3%的平均精度,相比原始YOLOv5模型,精确度提高了14%,召回率提高了10.2%,平均精度提高了12.1%,上述结果表明,改进后的算法可有效检测出肺结节。
To address the challenges of detecting small nodules in lung CT images and achieving a balance between lightweight and high-precision with the existing lung nodule detection algorithms,a high-precision and lightweight lung nodule detection algorithm based on improved YOLOv5 is proposed.The main improvements are focused on 4 aspects.(1)Replacing the stride-2 downsampling operation in the YOLOv5 backbone with space-to-depth downsampling operations to enhance fine feature extraction for detecting small nodules more comprehensively.(2)Employing an asymptotic feature pyramid network in the YOLOv5 neck to establish connections among feature maps from different paths,thereby enhancing interaction among different hierarchical levels.(3)Introducing global context-aware attention to the end of YOLOv5 neck network for improving the model's ability to understand key features and semantic information of lung nodules from a global perspective.(4)Utilizing the loss rank mining approach to strategically train on hard samples,thereby strengthening the model's discrimination ability.The improved algorithm achieves 96.0%precision,95.0%recall rate and 97.3%average precision on the LUNA16 dataset,which are 14.0%,10.2%and 12.1%higher than the original YOLOv5 model,demonstrating its effectiveness for lung nodule detection.