改进集成技术在甲状腺超声图像分类中的应用研究

发布时间：2018-11-10 09:33

【摘要】：随着超声成像和医学诊断技术的发展，超声已成为甲状腺癌检查的主要手段之一。当前，甲状腺癌的确定主要通过医生对B超图像的定性判别来完成，但由于甲状腺癌生物学特性多变以及各大医院诊断侧重点的不一，使得诊断结果易受医生的经验、水平、状态等因素的影响，诊断结果的准确度很难保证。因此，需要建立一种客观的方法，为医师诊断甲状腺疾病提供必要的辅助手段。集成学习是机器学习中的一种新型技术，它是在对新的实例进行分类的时候，把若干个体分类器集成起来，通过对多个分类器的分类结果进行某种组合来决定最终的分类，通常情况下，集成学习能够获得比单个分类器更好的性能。论文将动态集成技术引入到医学图像的分类问题中，研究如何利用动态集成算法在分类上的优势解决甲状腺B超图像分类识别中识别率低的问题。针对上述存在的问题，本文对特征提取量化、集成算法等方面进行了深入的研究，主要获得以下研究成果： 1.针对传统集成算法无法获得稳定分类精度的问题，对基于聚类的动态集成算法进行了改进，改进k-means聚类算法的目标函数和与能力区域距离计算公式，得到一种新的聚类和距离测量标准，提高了集成算法的分类准确率；同时，，提出了一种选择加权动态集成方法，采用多个分类器进行并联集成，以此来增加分类模型的稳定性；最后，通过实验证明了本文改进算法的有效性。 2.通过分析甲状腺B超图像，研究甲状腺良恶性结节在超声图像上的不同特点，综合考虑临床鉴别甲状腺结节的各种特征，分别对其进行量化，并提出了针对甲状腺结节特有的微钙化度度量方法，最终提取了最能描述结节性质的圆形度、衰减系数、微钙化度等9个特征参数作为甲状腺疾病数据集，为医生提供较为客观的量化参数。 3.为衡量本文算法分类器的性能，将本文基于改进动态集成的方法与相似研究常用的线性判别、BP神经网络和SVM算法进行了分类效果的比较，证明了本文算法的优势。
[Abstract]:With the development of ultrasound imaging and medical diagnosis, ultrasound has become one of the main methods of thyroid cancer examination. At present, the determination of thyroid cancer is mainly accomplished by the qualitative identification of B-ultrasound images by doctors. However, because of the changeable biological characteristics of thyroid cancer and the different diagnostic emphases in major hospitals, the diagnosis results are easily subject to the doctor's experience and level. The accuracy of diagnostic results is difficult to ensure due to the influence of state and other factors. Therefore, it is necessary to establish an objective method for physicians to diagnose thyroid diseases. Ensemble learning is a new technology in machine learning. When classifying new examples, it integrates several individual classifiers and determines the final classification by combining the classification results of multiple classifiers. In general, ensemble learning can achieve better performance than a single classifier. In this paper, the dynamic integration technique is introduced into the medical image classification problem, and how to use the dynamic integration algorithm to solve the problem of low recognition rate in the classification and recognition of thyroid B ultrasound image is studied. In view of the above problems, this paper makes a deep research on the feature extraction and quantization, integration algorithm and so on. The main research results are as follows: 1. Aiming at the problem that the traditional ensemble algorithm can not obtain the stable classification accuracy, the dynamic ensemble algorithm based on clustering is improved to improve the objective function of the k-means clustering algorithm and the calculation formula of distance between the clustering algorithm and the capability region. A new clustering and distance measurement standard is obtained, which improves the classification accuracy of the ensemble algorithm. At the same time, a selective weighted dynamic ensemble method is proposed, in which multiple classifiers are used for parallel integration to increase the stability of the classification model. Finally, the effectiveness of the improved algorithm is proved by experiments. 2. By analyzing the B-ultrasound images of thyroid gland, the different characteristics of benign and malignant thyroid nodules on ultrasound images were studied, and the clinical features of distinguishing thyroid nodules were comprehensively considered and quantified respectively. Finally, nine characteristic parameters, such as roundness, attenuation coefficient and microcalcification degree, which can best describe the nature of thyroid nodules, are extracted as data sets of thyroid diseases. To provide more objective quantitative parameters for doctors. 3. In order to evaluate the performance of the classifier in this paper, the improved dynamic ensemble method is compared with the linear discriminant, BP neural network and SVM algorithm, and the advantages of this algorithm are proved.
【学位授予单位】：河北大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：R445.1;R736.1;TP391.41

【参考文献】