改进密集连接网络的胸部多疾病X光图像分类算法

Improved chest multi-disease X-ray image classification algorithm for densely connected networks

为改变高度依赖人工读胸片的传统诊疗方式,关注疾病之间关联性,实现多病种自动识别,缓解医学影像诊断压力,减少误诊和漏诊,针对胸部疾病多标签分类提出了DS-EANet121模型.在DenseNet121网络的基础上采用动态激活函数,使变化的输入不断适应网络;引用SoftPool最大程度保留特征信息,并且为关注到更多的局部特征,融合ECA注意力机制,在适当跨信道交互的同时提升模型的性能.最终得到平均AUC为0.897,平均ACC为0.842.实验结果表明:改进后的DS-EANet121较原始网络分类精度上有明显提升,有一定的临床应用价值.

In order to change the traditional diagnosis and treatment methods that are highly dependent on manual reading of chest radiographs,the correlation between diseases was amalyzed to identify multipke diseases automatially.This can relieve the pressure of medical imaging diagnosis,and reduce misdiagnosis and missed diagnosis.The DS-EANet121 model was proposed for the multi-label classification of chest diseases,and the dynamic activation function was used on the basis of the DenseNet121 network to continuously adapt the changing inputs to the network.SoftPool was refered to maximize the retention of feature information.More local features was considered,the ECA attention module was integrated to improve the performance of the model while appropriating cross-channel interaction.The average AUC was 0.897 and the average ACC was 0.842.The experimental results show that the improved DS-EANet121 has obviously improved the classification accuracy compared with the original network,and has certain clinical application value.