基于细粒度的肺炎识别方法

Fine-Grained Pneumonia Recognition Method

【目的】针对肺炎识别案例中存在的数据集数量分配不均、数量少、类别间差异小等问题,提出一种基于卷积神经网络判别模块的肺炎识别方法。【方法】首先,将网络骨干设定为预训练的121层DenseNet网络,并冻结相关参数,以迁移学习的方式来解决数据量少的问题,再将网络的中间层定义一组额外的卷积滤波器,通过学习这组滤波器,可以捕获类别之间的高信号区域,而不需要额外的边框注释,以此解决肺炎识别中的类间样本差异小的问题。最后定义网络的损失函数为FocalLoss以解决数据集数量分配不均的问题。【结果】用提出的新方法在Chest-X-Ray Image数据集上进行实验。实验结果显示:该方法的准确率达到了95%,比传统的迁移学习准确率提升了10%,比未进行迁移学习的轻量级卷积神经网络的准确率则普遍提升了20%。【结论】该方法能够针对肺部X光片做出是否患有肺炎的判断,且定位出感染区域。

[Purposes]Aiming at the problems of uneven distribution of data sets,small numbers,and small differences between categories in pneumonia identification cases.[Methods]A pneumonia identification method is proposed based on the convolutional neural network discriminant module.First,the backbone of the network is a pre-trained 121-layer DenseNet network,and the relevant parameters are frozen to solve the problem of low data volume by means of transfer learning,and then a set of additional convolution filters are defined in the middle layer of the network.The group filter can capture the high signal area between the categories without the need for additional border annotations,so as to solve the problem of small sample differences between the categories in pneumonia recognition;finally define the loss function of the network as FocalLoss to solve the number of data sets of the problem of uneven distribution.[Findings]The proposed method is tested on the Chest-X-Ray Image dataset.The experimental results show that the accuracy of the method has reached 95%,which is 10%higher than the traditional transfer learning,and is lighter than the non-transfer learning.Convolutional neural networks have generally increased by 20%.[Conclusions]This method can determine whether there is pneumonia based on lung X-rays and locate the infected area.