基于超声和钼靶的两阶段乳腺癌诊断系统

Two-stage breast cancer diagnosis system based on ultrasound and mammogram images

摘要:

全球范围内乳腺癌发病率持续上升,乳腺癌的异质性导致良恶肿瘤在超声声像上呈现不同程度交叉和重叠,只利用一种图像信息不能得到满意的分类效果.因此本文提出基于超声和钼靶的两阶段乳腺癌诊断系统:第一阶段用拒绝分类方法对乳腺超声图像进行识别,对置信度高的肿瘤赋予其类别标签,对不确定的肿瘤不予分类；第二阶段对这些没有分类的肿瘤使用钼靶图像进行辨别.该系统利用多模态图像信息,筛查难以辨识的超声图像,辅以钼靶图像信息,对乳腺肿瘤进行诊断.本研究采用的超声和钼靶图像数据由哈尔滨医科大学肿瘤医院和哈尔滨医科大学第一附属医院提供,分别对拒绝分类方法和两阶段诊断系统进行实验验证和方法对比.与只使用乳腺超声特征的诊断系统相比,本文提出的两阶段诊断系统性能更优:正确率为92.59%,AUC为0.933 3,G-mean为93.09%、敏感性为86.67%,特异性为100%,阳性预测值为100%,阴性预测值为85.71%,马修斯相关系数为0.861 9.实验结果表明,增加钼靶图像信息可提高只使用超声图像的单一模态诊断系统的性能.

Abstract:

The incidence of breast cancer continues to rise worldwide. Due to the heterogeneity of breast cancer, there is an overlap between benign and malignant tumor images that only using a type of images cannot obtain satisfactory classification results. This paper proposes a two-stage breast cancer diagnosis system based on ultrasound and mammogram images. In the first phase, an abstaining classification method is used to classify breast ultrasound (BUS) images, in which some BUS tumors are classified with high confidence, and the uncertain tumors are not classified. These unclassified tumors are then classified using mammogram images in the second stage. Supplemented by mammogram information, the system can diagnose breast cancer by utilizing multimodal image information to screen for unrecognizable ultrasound images. The ultrasound and mammogram images used in this study were provided by the Cancer Hospital of Harbin Medical University and the First Affiliated Hospital of Harbin Medical University. The abstaining method and the two-stage diagnosis system were validated in experiments. Compared with diagnostic systems using only BUS features, the proposed two-stage diagnostic system provided better performance with the accuracy of 92.59%, AUC of 0.933 3, G-mean of 93.09%, sensitivity of 86.67%, specificity of 100%, positive predictive value of 100%, negative predictive value of 85.71%, and Matthew’s correlation coefficient of 0.861 9. Experimental results demonstrate that adding mammogram information can increase the performance of the diagnosis system that uses BUS images only.