针对目前基于深度学习的肺结节检测算法中不同深度与尺寸的特征信息间没有相互交流的问题,提出了一种基于密集残差连接的肺结节检测模型.本模型在 3D U-Net网络的基础上引入密集连接,充分利用网络中肺结节特征图,实现不同层的特征信息的结合,提高结节特征的利用率;同时结合残差结构,避免了网络加深后出现的梯度消失问题;引入通道注意力机制,对不同通道的结节特征赋予权重,提高结节的识别率;在 3D U-Net网络的编码解码部分间的跳跃连接中使用转置卷积,融合不同尺度与不同深度的特征.所提算法在肺结节公共数据集LUNA16 上进行十折交叉验证,以无限制受试者操作特征为评价指标,实验结果表明,在假阳率为 0.125、0.25、0.5、1、2、4、8 这 7 个点上,平均敏感度为 0.852,相较于基准模型提升5.5%.所提出的肺结节检测算法相比基准模型提高了检测敏感度,较好的实现对肺结节的检测.
Lung Nodule Detection Method Based on Dense Residual Connection
Aiming at the problem that the feature information of different depths and sizes does not communicate with each other in the current deep learning based lung nodule detection algorithm,a lung nodule detection model based on dense connection and residual con-nection is proposed.This model introduces dense connections on the basis of 3D U-Net network,which makes full use of lung nodule features in the network to realize the combination of feature information of different layers,and can improve the utilization rate of nodule features.By combining with the residual structure,the problem of gradient disappearance after network deepening is avoided.The chan-nel attention mechanism is introduced to produce weights to the nodule features of different channels to improve the nodule recognition rate.The transposed convolution is used in the skip connection between the encoding and decoding parts of 3D U-Net network,and fea-tures of different scales and depths are integrated at the same time.The proposed algorithm is performed by ten-fold cross-validation on the lung nodule public dataset of LUNA16 and evaluated in terms of free-response receiver operating characteristic.The experimental results show that the average sensitivity is 0.852 at the seven points where the false positive rate is 0.125,0.25,0.5,1,2,4 and 8,which is 5.5%higher than the benchmark model.Compared with the benchmark model,the proposed lung nodule detection algorithm improves the detection sensitivity and better realizes the detection of lung nodules.
万方数据
维普
