基于放射组学的肺ROI特征提取与选择和结节的良恶性分类

发布时间：2018-05-05 13:08

本文选题：CT图像 + 放射组学　；参考：《河北大学》2017年硕士论文

【摘要】：众所周知,癌症是目前全球发病率和死亡率最高的疾病之一。通过活体检查的方式对肿瘤的良恶性进行诊断,不仅需要对患者做入侵手术,而且不利于医生对肿瘤异质性的观察。放射组学通过对放射影像中获取的可描述性特征的量化和分析,将肿瘤表型和疗效评估对应起来,为医生对肺癌患者进行临床上的诊断提供了良好的依据。近年来,深度学习方法被广泛应用于医学图像处理领域。利用原始放射影像进行深度学习虽然可以获得良好的识别效果,但由于其封闭的学习方式,无法得知图像特征与分类结果间的关系,而采用传统的浅层机器学习方法通常具有一定的局限性。本文主要主要研究工作如下:(1)采用传统的统计方法从肺部CT图像的感兴趣区域中提取了几何特征、纹理特征和直方图特征,共143维特征向量作为原始特征集,用于结节的良恶性分类。(2)受基因组学在基因选择中的启发,在Relief特征选择算法中引入递归特征排除策略,形成RFE-Relief特征选择算法,克服了Relief算法中不能去除冗余特征的缺点,最终得到包含有46维特征向量且与结节良恶性相关性强的低维特征子集。(3)构建了一个由三层受限玻尔兹曼机构成的深度置信网络模型,并在顶层加入Softmax分类器,将由46维特征向量构成的特征子集作为DBN的输入,通过对RBM和Softmax的逐层训练和微调实现对结节的良恶性分类。实验结果表明,最终的分类精确度为93.8%。通过特征选择,不仅降低了算法的运行时间和效率,还提高了分类的精确度,可以辅助临床医生进行肺癌诊断。本文最后还对DBN模型在结节良恶性分类中的隐含层节点数和隐含层数进行了讨论,结果显示本文构建的DBN模型可以较好的完成结节良恶性的分类任务。
[Abstract]:It is well known that cancer is one of the highest morbidity and mortality in the world. The diagnosis of benign and malignant tumors by living examination not only requires invasive surgery, but also is not conducive to doctors' observation of tumor heterogeneity. By quantifying and analyzing the descriptive features obtained from radiographic images, radiology corresponds the tumor phenotype to the evaluation of curative effect, which provides a good basis for the clinical diagnosis of lung cancer patients. In recent years, depth learning has been widely used in the field of medical image processing. Although the deep learning of the original radiographic image can obtain good recognition effect, because of its closed learning mode, it is impossible to know the relationship between the image features and the classification results. However, the traditional shallow machine learning method usually has some limitations. The main work of this paper is as follows: (1) the geometric feature, texture feature and histogram feature are extracted from the region of interest of lung CT image by traditional statistical method. A total of 143 dimensional feature vector is used as the original feature set. Inspired by genomics in gene selection, recursive feature exclusion strategy is introduced into Relief feature selection algorithm to form RFE-Relief feature selection algorithm, which overcomes the disadvantage that redundant features can not be removed in Relief algorithm. Finally, a low dimensional feature subset containing 46 dimensional feature vectors with strong correlation with benign and malignant nodules is obtained. A depth confidence network model based on a three layer constrained Boltzmann mechanism is constructed, and a Softmax classifier is added to the top layer. A feature subset composed of 46 dimensional feature vectors is used as the input of DBN. The classification of benign and malignant nodules is realized by training and fine-tuning the RBM and Softmax layer by layer. The experimental results show that the final classification accuracy is 93. 8%. Feature selection not only reduces the running time and efficiency of the algorithm, but also improves the accuracy of classification, which can assist clinicians in the diagnosis of lung cancer. Finally, the number of hidden nodes and the number of hidden layers of DBN model in the classification of benign and malignant nodules are discussed. The results show that the DBN model constructed in this paper can accomplish the classification of benign and malignant nodules.
【学位授予单位】：河北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R734.2;TP391.41

【参考文献】