基于局部感受野和半监督深度自编码的肺结节检测方法

摘要

Deep learning in pulmonary imaging research mainly concentrated on CT image.And the key step of pulmonary disease treatment are how to make pulmonary nodules detection fast and exactly.The detection of pulmonary is a challenging task,and existing research is difficult to get a higher rate.For this problem,an improved deep sparse autoencoder pulmonary nodule recognized method is proposed.First,analysis local multi layer features.Second,semi-supervised sparse autoencoder is used to automatically obtain the nodular features in lung images.Finally,achieve accurate identification of pulmonary nodules by integrating a variety of clinical information.The experimental results show that the method can achieve accuracy 90.14％,sensitivity 89.67％,and average recognition rate 96.64％.It's better than other methods of identification performance and more suitable for accurate identification of pulmonary nodules.%深度学习在肺部影像方面的研究主要集中于肺部CT图像.对肺结节的快速准确检测是肺部疾病治疗的关键步骤.结节检测本身就是一项具有挑战性的工作,且已有的研究均很难得到较高的检测率.针对这样的问题,提出一种改进的深度半监督稀疏自编码的肺结节检测方法.首先,采用局部感受野对肺结节图像进行多层特征提取.然后,利用半监督稀疏自编码自主学习肺部影像中的结节特征.最后,融合多种临床信息实现对肺结节的准确检测.实验结果表明,该方法可以达到准确率90.14％,敏感度89.67％和平均检测率96.64％,明显优于其他方法检测性能,更适用于肺结节的精准检测.