摘要
传统肺结节良恶性诊断中,难以提取具有显著性和区分性的结节特征,会造成结节诊断准确率低、误诊率高等问题,为此提出一种基于三维尺度不变特征的肺结节良恶性诊断方法。综合考虑临床结节的三维特征和局部细节特征,选取结节序列图像的尺度不变特征点,联合3DSIFT描述子和灰度差累加直方图,统计获得每个特征点的尺度不变描述子,充分表征病灶的局部细节,利用DFCM聚类算法和BoW模型计算得到结节的特征表示,训练SVM完成结节的良恶性诊断。实验结果表明,该方法实现了优异的诊断效果,与不同文献中方法的对比中,有明显优势,且恶性度为3的结节更类似于良性结节。
In the traditional method of benign and malignant diagnosis of pulmonary nodules,it is difficult to extract the nodular features with significant and discriminating characteristics,resulting in low diagnostic accuracy of nodules and high misdiagnosis rate.A diagnostic method for benign and malignant pulmonary nodules based on three-dimensional scale invariant features was then proposed.Considering the three-dimensional features and local details of clinical nodules,scale invariant feature points of nodule sequence images were selected.Combining 3D SIFT descriptor and gray scale difference cumulative histogram,the scale invariant descriptor of each feature point was obtained to fully characterize the local details of the lesion.The feature representation of the nodule was calculated using DFCM clustering algorithm and BoW model.SVM was trained to diagnose the benign and malignant nodules.Experimental results show that,the proposed method achieves excellent diagnostic results.Compared with the methods in different literature,this method has obvious advantages,and it is found that nodules with malignancy of 3 are more similar to benign nodules.
作者
吴化禹
强彦
王三虎
刘希靖
原杰
WU Hua-yu;QIANG Yan;WANG San-hu;LIU Xi-jing;YUAN Jie(College of Computer Science and Technology,Taiyuan University of Technology,Jinzhong 030600,China;Department of Computer Science and Technology,Lvliang University,Lvliang 033000,China;School of Software,Shanxi Agricultural University,Jinzhong 030801,China;PET/CT Center,Shanxi Provincial People’s Hospital,Taiyuan 030024,China)
出处
《计算机工程与设计》
北大核心
2019年第10期2843-2848,共6页
Computer Engineering and Design
基金
国家自然科学基金项目(61572344)
虚拟现实技术与系统国家重点实验室开放基金项目(BUAA-VR-17KF-14)
山西省回国留学人员科研基金项目(2016-038)
关键词
肺结节
序列图像
尺度不变特征
三维
良恶性诊断
pulmonary nodules
sequence images
scale-invariant features
three-dimensional
benign and malignant diagnosis
