【作者】 唐盛；

【导师】 段会龙；

【作者基本信息】 浙江大学 ， 生物医学工程， 2009， 博士

【摘要】 医学超声图像识别研究始终受到超声图像存在斑点噪声、信噪比低、边缘信息丢失等因素的影响。本文首先针对超声图像识别问题的固有特点，对不均衡类样本问题和数据噪声问题进行研究，提出了一种特征数据预处理机制，并结合多分类器融合提出了一种超声子宫图像识别框架。然后设计并实现了一种快速全自动的超声子宫图像节育环物体识别及定位算法。最后再设计并实现了一种快速全自动的超声图像胎儿性别部位识别及定位算法。本文研究内容依次包括以下几个方面：1、研究了不均衡类样本问题，概述了该领域的发展现状，并针对人造少数类样本生成机制进行研究，提出了一种人造样本过采样算法：ADOMS算法。基于UCI特征集开展实验验证工作，实验结果表明与其他相关算法相比，所提ADOMS算法可以更为有效地减轻不均衡类样本环境下分类器分类性能下降情况。2、研究了数据噪声问题，概述了该领域的发展现状，并提出了一类基于数学形态学思想的形态学去噪算法。基于UCI特征集开展实验验证工作，实验结果表明与其他相关算法相比，所提形态学去噪算法可以更为有效地改善分类器分类性能。3、提出了一种同时处理不均衡类样本问题和数据噪声问题的串联处理机制，通过实验验证了该串联处理机制在超声子宫图像识别问题上的良好效果。随后结合特征降维进一步提出了一种特征数据预处理机制，并在此基础上结合多分类器融合提出了一种超声子宫图像识别框架。4、设计并实现了一种快速全自动的超声子宫图像节育环物体识别及定位算法，其包括全自动超声子宫图像分割和特定模式识别框架两大部分，通过感兴趣物体(Obiect of interest，OOI)进行连接。基于由719幅超声子宫图像所组成实验图像库开展算法性能验证工作，实验结果表明所提算法性能优秀，其处理过程完全自动，在普通PC机上平均耗时仅527ms／幅，有环图像识别及定位准确率达81．1％，无环图像识别准确率达94．7％。5、设计并实现了一种快速全自动的超声图像胎儿性别部位识别及定位算法，其可分为粗分类阶段和细分类阶段两大部分，通过感兴趣像素(Point of interest，POI)进行连接。基于由658幅阳性图像(图中含胎儿性别部位)和500幅阴性图像(图中不含胎儿性别部位)所组成实验图像库开展算法性能验证工作，实验结果表明所提算法性能优秀，其处理过程完全自动，在普通PC机上平均耗时仅为453ms／幅，阳性图像识别及定位准确率达80．9％，阴性图像识别准确率达83．8％。更多还原

【Abstract】 The research development of medical ultrasound image recognition is still cumbered by the poor quality of ultrasound image, including speckle, low signal-to-noise ratio, lost of edge information, etc. In this dissertation, focusing on the instinctive characteristics of ultrasound image recognition problem, the class imbalance problem and the data noise problem are studied at first, and then an ultrasound uterus image recognition framework is proposed, which is based on the proposed feature data preprocessing mechanism and the multiple classifiers fusion. After that, a fast and automatic ultrasound uterus image recognition algorithm for the intra-uterine device is proposed and realized. At last, a fast and automatic ultrasound image recognition algorithm for fetal genital organ is proposed and realized.The main researches of this dissertation are listed as following:1, The related work of dealing with the class imbalance problem is firstly reviewed, and then a proper generation mechanism of synthetic minority class examples is discussed. According to the analysis, a novel oversampling algorithm with synthetic examples, ADOMS, is proposed. The experiments are arranged on the UCI datasets and the experimental results show that comparing with other relative methods, algorithm ADOMS is able to alleviate the deterioration of the classification performance of the classifiers effectively.2, The related work of dealing with the data noise problem is firstly reviewed, and then based on the concept of mathematic morphology, a series of morphological data cleansing algorithms are proposed. The experiments are arranged on the UCI datasets and the experimental results show that these morphological data cleansing algorithms can effectively improve the classification performance of the classifiers, comparing with other relative methods.3, A concatenation mechanism is proposed to handle the class imbalance problem and the data noise problem together, and the effect of the proposed concatenation mechanism to the ultrasound uterus image recognition problem is confirmed by the experiments. And then, combing a proposed feature data preprocessing mechanism and the multiple classifiers fusion, an ultrasound uterus image recognition framework is proposed.4, A fast and automatic ultrasound uterus image recognition algorithm for the intra-uterine device (IUD) is proposed and realized. The algorithm is composed of automatic ultrasound uterus image segmentation and the specific recognition framework, which are connected by object of interest (OOI). Based on 719 ultrasound uterus images, the experiments are carried out. The experimental results show that the proposed algorithm is fully automatic, and the average time-consuming is 527 milliseconds per frame, as well as the accuracy for the uterus image with IUD is 81.1% and the accuracy for the uterus image without IUD is 94.7%.5, A fast and automatic ultrasound image recognition algorithm for fetal genital organ is proposed and realized. The algorithm is composed of rough classification stage and fine classification stage, which are connected by pixel of interest (POI). Based on 658 positive images (the ultrasound images which containing fetal genital organ) and 500 negative images (the ultrasound images without fetal genital organ), the experiments are carried out. The experimental results show that the proposed algorithm is fully automatic, and the average time-consuming is 453 milliseconds per frame, as well as the accuracy for the positive images is 80.9%, and the accuracy for the negative images is 83.8%.更多还原