基于双视角和多分类器信息融合的乳腺钼靶图像肿块分类研究

摘要

乳腺肿块良恶性分类是计算机辅助诊断(CAD)的重要环节,如何提高分类的正确率和稳定性是分类研究的重点.本研究提出了4种基于双视角和多分类器信息融合的乳腺钼靶图像肿块分类模式.其中,模式1是单视角下的多分类器融合;模式2是分别先对每个分类器在两个视角下的输出进行视角融合,再对其融合结果进行多分类器融合;模式3是分别先在每个视角下进行多分类器融合,再对两个视角的多分类器结果进行视角融合;模式4是先对特征向量在两个视角下的取值进行融合,再基于新的特征向量进行单分类器分类和多分类器融合.从南佛罗里达大学DDSM数据库中随机选择的148个良性肿块和148个恶性肿块,对这4种分类模式的效果进行比较.实验结果表明,在肿块分类的正确率、敏感性、特异性和稳定性等方面,模式2和模式3的表现均优于模式1和模式4.%Classification of masses in mammography is an important part of Computer-Aided Diagnosis (CAD).How to improve the accuracy and stability of the classification is the focus of the current studies.Based on information fusions of multi-view and multi-classifier, four classification models on masses in mammograms are proposed.The first classification model uses multi-classifier fusion in single-view; In the second model, the outputs of each classifier in two views are fused, and then these results of multi-view are used for multiclassifier fusion; In the third model, fusion of multi-classifier is applied in each view, then the two fusion results are used for multi-view fusion; In the fourth model, feature vectors of two views are fused, then classification of singe-classifier and multi-classifier fusion are used.In the experiments, we randomly selected 148 benign masses and 148 malignant masses from the DDSM database.The experiment results revealed that the second and third models are superior to the first and fourth models in terms of accuracy, sensitivity, specificity and stability.