马巧梅, 梁昊然, 王明俊, 程鑫. 多策略融合的肺结节检测模型与算法[J]. 微电子学与计算机, 2021, 38(6): 38-44.
引用本文: 马巧梅, 梁昊然, 王明俊, 程鑫. 多策略融合的肺结节检测模型与算法[J]. 微电子学与计算机, 2021, 38(6): 38-44.
MA Qiao-mei, LIANG Hao-ran, WANG Ming-jun, CHENG Xin. Multi-strategy fusion model and algorithm for lung nodule detection[J]. Microelectronics & Computer, 2021, 38(6): 38-44.
Citation: MA Qiao-mei, LIANG Hao-ran, WANG Ming-jun, CHENG Xin. Multi-strategy fusion model and algorithm for lung nodule detection[J]. Microelectronics & Computer, 2021, 38(6): 38-44.

多策略融合的肺结节检测模型与算法

Multi-strategy fusion model and algorithm for lung nodule detection

  • 摘要: 肺结节早期检测可以提高病人的生存率,自动检测算法可以有效辅助医生进行诊断.为了提高肺结节检测精确度、降低漏诊率,提出了基于循环残差注意力门机制的U-Net(Recurrent Residual Attention Gate U-Net,R2AGU-Net)肺结节检测模型.首先在原始的U-Net基础上改进,添加循环残差卷积模块并融合注意力门机制,在增强特征提取性能的同时将注意力放在目标结节区域,通过抑制无关的特征响应获得较高的检测精度;其次改进损失函数解决肺结节图像数据不均衡问题,获得较高的检测敏感度;最后通过三维卷积神经网络(3D CNN)分类候选结节,降低检测的假阳性.在两个数据集上进行实验验证,结果表明本文提出的算法提升了检测速度和敏感度,取得了比现有算法更好的性能,具有较好的泛化能力。

     

    Abstract: Early detection of lung nodules can effectively improve the survival rate of patients. Automatic detection algorithm can effectively assist doctors in diagnosis. In the lung nodule detection task, in order to improve the detection accuracy and reduce the missed diagnosis rate, the lung nodule detection algorithm based on the recurrent residual attention gate U-Net (R2AGU-Net) was proposed. Firstly, recurrent residual convolution module and fusion attention gate mechanism were added in U-Net. The improved method can enhance the feature extraction performance, focus on the target nodule area and suppress irrelevant feature responses to obtain higher detection accuracy. Secondly, improved loss function can solve the problem of unbalanced image data of lung nodules to obtain higher detection sensitivity. Finally, the candidate nodules are classified by a three-dimensional convolutional neural network (3D CNN) to reduce false positive. Experimental verification on two datasets shows that the algorithm proposed in this paper improves the detection speed and sensitivity, achieves better performance than existing algorithms, and has better generalization capabilities.

     

/

返回文章
返回