结合层次法与主成分分析特征变换的宫颈细胞识别

对宫颈细胞进行多分类可以自动识别出不同状态的细胞，进而为宫颈癌诊断提供科学依据。在用6种多分类算法进行实验后，选取支持向量机作为基分类器，先用一对一策略训练6个分类器进行3分类，然后再训练1个2分类器，这种二层4分类方法提高了识别准确率。考虑不同层特征模式的差异性，在保证识别性能的同时，每层分类前先采用主成分分析法将原始154维特征变换到低维空间，去除冗余特征，加快识别速度。实验证明，所提层次主成分分析法在宫颈细胞分类中相比6种传统多分类方法有更高的识别准确率，可达90%以上；识别速度也较普通层次法提升了21.31%。
In order to recognize multi-class cervical cells automatically, a hierarchical method with PCA (principal component analysis) feature transformation was proposed and this cell recognition could provide the evidence for cervical cancer diagnosis. The cervical cell recognition was treated as a 4-class classification problem. There were two levels in this hierarchical method. First, one-versus-one strategy was used to train 6 SVM (support vector machine) classifiers to do a 3-class classification. Second, abnormal cells in one type of 3 categories were classified by a 2-class SVM. To optimize the feature combination and reduce the running time, a feature transformation method named PCA was adopted to transform the original feature vector into low dimension feature space. The experiments show that the proposed hierarchical PCA recognition method is faster than the common hierarchical method at a ratio of 21.31%, and can distinguish 4 cervical cell categories better than 6 other traditional methods and achieve above 90% accuracy.