摘要
在过去的几年中,肺癌是癌症相关死亡的主要原因。提出一种针对低剂量计算机断层扫描(CT)影像精细化预处理条件下的SE-CapsNet分类方法,解决传统肺结节诊断方法中分类精度低、假阳性高等问题。改进胶囊神经网络分类算法:对最新Hinton的胶囊神经网络进行改进,引入新的非线性激活向量,避免全局向量压缩;采用特征重标定的方法,在特征通道层面进行模型优化。在标定的感兴趣区域,利用自动阈值法对CT影像进行预处理,并在中心结节处进行样本采样,获得预处理结果数据样本。选用内含1 010个病例的公开数据集LIDC-IDRI和某医院30个脱敏肿瘤患者病例,评估改进的SE-CapsNet算法,评价指标包括准确性、敏感性和特异性。在LIDC-IDRI数据集与医院数据集中,SE-CapsNet算法的平均准确率分别达到95.83%和94.67%,优于基于Caps Net分类算法的平均准确率。此外,在分类算法的耗时方面也具有明显优势,改进的胶囊网络能够更快地收敛,得到稳定的结果。
Over the past few years,lung cancer has been the leading cause of cancer-related deaths. This paper proposed a SE-CapsNet classification method for the low-dose computed tomography(CT)image refinement preprocessing conditions. Our work solved the problems of low classification accuracy and high false positives in traditional lung nodule diagnosis methods,which improved the capsule neural network classification algorithm,including improving the latest Hinton’s capsule neural network,introducing new non-linear activation vectors,avoiding global vector compression,and optimizing the model at the feature channel level by feature reweight.We used the automatic threshold method to process the CT images by calibrating the region of interest,and took the samples at the central nodule to obtain data samples of the pre-processing results. The public data set LIDCIDRI containing 1010 cases and 30 cases of desensitized tumor patients eliminated sensitive information from hospital were used to evaluate the improved SE-CapsNet algorithm. The evaluation criteria mainly included accuracy,sensitivity and specificity. In the LIDC-IDRI dataset and the hospital dataset,the average accuracy of the SE-CapsNet algorithm reached 95. 83% and 94. 67%,respectively,which was superior to that by CapsNet classification algorithm. In addition,the classification algorithm also had obvious advantages in terms of time consumption,and the improved capsule network converged faster to obtain stable results.
作者
叶枫
王路遥
洪卫
丁国军
车镓荣
Ye Feng;Wang Luyao;Hong Wei;Ding Guojun;Che Jiarong(School of Management,Zhejiang University of Technology,Hangzhou 310023,China;Department of Thoracic Oncology,Zhejiang Cancer Hospital,Hangzhou 310023,China;Department of Radiology,Zhejiang Cancer Hospital,Hangzhou 310023,China)
出处
《中国生物医学工程学报》
CAS
CSCD
北大核心
2021年第1期71-80,共10页
Chinese Journal of Biomedical Engineering
基金
国家社会科学基金(18BJY148)。
