基于改进LSTM的肺结节良恶性预测方法研究

Benign and Malignant Prediction of Pulmonary Lesions based on improved Long Short-Term Memory network

目前深度学习计算机辅助诊断技术已经相对成熟,可以实现对肺部医学影像中可疑结节的识别,甚至可以完成对结节大小、类型等信息的精确测量,为医师临床诊断肺结节的良恶性提供有利依据。然而,传统常规的深度学习网络构建过程中使用的均为单独孤立的肺结节影像片段,并没有从时间序列的角度上对特征进行研究和探索。文章提出了一种基于自编码和长短期记忆网络的肺结节良恶性预测算法。首先,利用自编码网络(SAE)自动提取出肺结节影像数据的深度隐藏特征。在此基础上,利用长短期记忆(LSTM)构建了一个双层构时间序列模型,学习了肺结节特征在时间序列上的变化情况。最后,在收集到的5760张随诊数据集上进行了对比验证,从实验结果中可以看出,该方法在对肺结节进行良恶性预测的精准度为91.51%,其识别精度和收敛性能均优于其他比较算法。

Toady.computer aided diagnosis technology with deep leaming has been relatively mature.It can not only realize the au-tomatic marking of suspicious nodules in CT images,but also complete the automatic measurement of nodule size,density and other information,so as to provide a favorable basis for doctors'clinical diagnosis.However,in the traditional deep leaming network con-struction process,only isolated lung nodule image fragments are used.The characteristics are not studied and explored from the perspective of time series.In this paper,an prediction algorithm for pulmonary nodules based on self-coding and long and short term memory network is proposed.Firstly,a auto-coding network(SAE)was used to automatically extract the depth hidden features of pulmonary nodule image data.On this basis,a two-layer time series model was constructed using deep learning to learn the changes of pulmonary nodule features in time series.Finally,we conducted experiments and evaluations on 5760 clinical data sets.The experimental results show that the accuracy of this method is 91.51%which is better than other comparison algorithms.