A method of breast masses diagnosis fused medical clinical feature
A method of breast masses diagnosis fused medical clinical feature
复制标题
DOI：
10.1109/fads.2017.8253214
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
中科院分区：
文献类型：
--
作者：
乳腺癌是肿瘤学中最常见的恶性肿瘤之一，严重威胁着女性的健康。多模态医学图像融合可以为临床医生提供更全面、准确的患者信息，提高乳腺癌诊断的准确性和治疗的成功率。临床医生在多模态医学图像中诊断乳腺肿瘤，不仅要看图像特征，还要关注患者的临床特征。临床特征对乳腺肿瘤的诊断很重要。但是，利用像素级影像学特征和多种临床特征而不经治疗对乳腺癌进行良恶性分类，会干扰分类结果。因此，提出了一种基于医学临床数据的特征融合方法，通过整合信息特征来提高质量分类率。该方法包括三个阶段。首先，低层图像特征映射到高层语义特征。然后采用基于典型相关分析的特征融合策略对不同语义特征组进行融合。分组的思想保护了具有不同呈现方式的特征的特定优点。最后，对特征融合产生的新特征向量采用加权集成策略进行分类。对103例患者109个乳腺病灶的实验结果表明，基于典型相关分析融合策略的分类方法的灵敏度、特异度和准确度分别为97.74%、95.43%和96.66%，优于传统的基于串联和并联特征融合的分类方法。
Breast cancer is one of the most common malignant cancers among oncology, which poses a serious health threat for females. Multimodal medical image fusion can provide clinicians with more comprehensive and accurate patient information, improving the accuracy of breast cancer diagnosis and treatment success rate. The clinician diagnoses the breast tumor in the multimodal medical image, not only look at the image features, but also pay attention to the clinical features of the patients. Clinical features are important for the diagnosis of breast tumors. However, the use of pixel-level imaging features and multiple clinical features without treatment for benign and malignant classification of breast cancer, will interfere with the classification results. Therefore, a novel feature fusion method based on medical clinical data is proposed to promote the mass classification rate by integrating informative features. The method consists of three stages. Firstly, low-level image features are mapped to high-level semantic features. Then, the feature fusion strategy based on canonical correlation analysis is used to fuse the different semantic feature groups. The idea of grouping protects the particular advantages of features that have different presenting ways. Finally, all of the new feature vectors generated by feature fusion are classified by weighted ensemble strategy. Experiments on 103 patients with 109 breast lesions have demonstrated that the proposed method based on canonical correlation analysis fusion strategy achieves sensitivity, specificity and accuracy of 97.74%, 95.43% and 96.66% respectively, which outperforms the traditional classification method based on series and parallel feature fusion.