基于乳腺X线肿块影像的计算机辅助诊断技术研究

发布时间：2018-03-29 09:28

本文选题：钼靶乳腺X线影像　切入点：计算机辅助诊断系统　出处：《浙江大学》2013年博士论文

【摘要】：乳腺癌是全球女性中最为频发的恶性肿瘤疾病和癌症死亡的首要原因。大量研究表明,早期诊断是扩大乳腺癌治疗方案的选择空间、降低乳腺癌患者死亡率的关键。钼靶乳腺X线摄影能显示触诊无法感知的早期病变,是乳腺癌早期诊断的首选方法。而基于乳腺X线影像的计算机辅助诊断(computer-aided diagnosis, CAD)技术能够协助放射科医生检测和分析可疑病灶,提高乳腺癌早期诊断的准确率。肿块是乳腺X线影像中的重要病理征象,其影像表现复杂多变,容易受到周围组织的干扰,是乳腺CAD系统的研究重点和难点之一。传统的乳腺CAD系统以病灶检测和病理分类为主要功能,其只提示可疑病变而不提供诊断依据的“黑箱”过程降低了医生对CAD系统的认可和信赖。近年来,一种基于图像内容检索(content-based image retrieval, CBIR)的CAD系统能够很好的克服传统CAD系统在辅助诊断方面的不足。本论文重点探讨了基于乳腺X线肿块影像的CBIR-CAD系统的关键技术,旨在提高肿块检测和辅助诊断的性能,为放射科医生提供有价值的“第二参考意见”。本论文的内容包括肿块分割、特征提取和优化、肿块亚型分类以及基于内容的相似肿块检索。 (1)肿块分割：提出了两种基于随机游走的自动肿块分割算法。第一种方法利用肿块等高线图自动标记随机游走算法所需的种子点,将背景种子点设置为包围肿块的闭合轮廓,抑制周围组织的影响。同时,利用等高线图的嵌套特性,不断扩张肿块种子点的范围,并利用随机游走算法获得一系列分割结果。最后,定义结合肿块尺寸、边界梯度和灰度信息的评价函数并选择最终的分割结果。该方法克服了半自动随机游走算法的应用局限性。第二种算法结合了随机游走和Chan-Vese (CV)活动轮廓模型的互补优势,首先利用等高线图自动标记种子点并进行初始随机游走分割。然后,利用随机游走获得的概率矩阵调制CV模型的能量函数,抑制轮廓泄露现象。同时,在轮廓演化过程中不断更新概率矩阵,并获得最终的分割结果。实验结果表明本论文提出的两种分割方法与几种现有的分割算法相比,具有更好的分割精度、适应性和鲁棒性。 (2)特征提取和优化：以乳腺影像报告和数据系统(Breast Imaging Reporting and Data System, BI-RADS)为依据,分别提取了肿块灰度、形状和边缘特征,并提出了6维基于肿块等高线拓扑变化的新特征。然后,综合五种特征性能评价指标,针对不同的分类目标,采用滤波式方法对上述多维特征进行优化。实验对比了优化前后的特征向量和分类性能,结果显示本论文提出的6维特征在区别肿块的5种边缘亚型方面表现出显著的优势。特征优化过程在降低特征维度的同时,提高了分类性能和效率。 (3)肿块亚型分类：提出了一种结合支持向量机(support vector machine, SVM)和二叉决策树(binary decision tree, BDT)的动态多分类方法,对肿块的4种形状亚型和5种边缘亚型进行分类。该方法结合了BDT较高的运算效率和SVM良好的分类性能,同时能够根据查询肿块的实际特征,动态调整节点分类器和分类步骤,从而有效抑制传统SVM-BDT方法的错误累积问题,在提高分类精度的同时,具有较高的训练和测试效率。 (4)基于内容的相似肿块检索：首先结合亚型分类器的概率输出结果和特征向量的欧式距离定义相似性测度函数,获得第一轮按例检索(query-by-example, QBE)结果。然后,在相关反馈(relevant feedback, RFb)环节,通过交互操作对第一轮检索结果进行相关性标记。针对实际应用中用户标定样本数量过少的问题,利用未标记样本与标记样本的核空间欧式距离进一步扩充了反馈样本的数量,并训练新的SVM相似性判断模型。最后,采用留一校验法(leave-one-out cross validation, LOOCV)和准确率召回率曲线(precision-recall curve, PRC)评价了QBE环节和RFb环节的检索性能。实验结果显示, QBE环节结合亚型分类器概率输出和欧式距离的相似性测度函数比单独使用欧式距离具有更好的检索性能。同时,在用户标定样本数和反馈次数一致的条件下,本论文的RFb环节对系统检索性能有更大的提升。
[Abstract]:Breast cancer is one of the most frequent malignant tumor diseases and cancer deaths among women in the world . A large number of studies have shown that early diagnosis is the key to expanding breast cancer treatment options and reducing the mortality of breast cancer patients . Mammography of mammography can show the early diagnosis of breast cancer . The computer - aided diagnosis ( CAD ) technique based on breast X - ray image can help the radiologist to detect and analyze suspicious lesions and improve the accuracy of early diagnosis of breast cancer .

In recent years , a CAD system based on content - based image retrieval ( CBIR ) can overcome the shortage of traditional CAD system in auxiliary diagnosis .

This paper focuses on the key techniques of CBIR - CAD system based on breast X - ray mass image , aiming at improving the performance of tumor detection and auxiliary diagnosis , providing valuable " second reference opinion " for radiologists . The contents of this paper include tumor segmentation , feature extraction and optimization , classification of tumor subtypes and retrieval of similar masses based on content .

( 1 ) Tumor segmentation : Two kinds of automatic mass segmentation algorithms based on random walk are presented . The first method uses the contour map to automatically mark the seed point needed by the random walk algorithm , and then sets the background seed point to surround the closed contour of the mass , and then obtains a series of segmentation results .

( 2 ) Feature extraction and optimization : Based on Breast Imaging Reporting and Data System ( BI - RADS ) , the gray scale , shape and edge characteristics of the masses are extracted respectively .

( 3 ) Classification of tumor subtypes : A dynamic multi - classification method combined with support vector machine ( SVM ) and binary decision tree ( BDT ) is proposed to classify the four types of tumor and five types of edge subtypes . The method combines the high operational efficiency of BDT and good classification performance of SVM , and can dynamically adjust the node classifier and classification step according to the actual characteristics of the masses , thereby effectively inhibiting the error accumulation problem of the traditional SVM - BDT method , and has higher training and testing efficiency while improving the classification accuracy .

( 4 ) Based on the similarity measure function of the content - based similarity measure , the results of query - by - example ( QBE ) are firstly combined with the probability output result of the subtype classifier and the Euclidean distance of the feature vector . Then , the retrieval performance of the QBE segment and RFb link is further expanded by interaction operation in the link of relevant feedback ( RFb ) . The results show that the similarity measure function of the probability output and the Euclidean distance of the combined subtype classifier in QBE segment has better retrieval performance than the Euclidean distance alone . At the same time , the RFb link of this paper has a greater improvement in the retrieval performance of the system under the condition of consistent user calibration sample number and feedback frequency .

【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：R816.4;R737.9

【参考文献】