Multiclass classification of vertebral pathologies with limited data

Inhalt:
For intelligent diagnosis of vertebral pathologies, the available labeled data are usually limited and the discrimination of multiple lesions is not obvious. To deal with this issue, a novel multiclass classification method with limited data is proposed, which helps assist in the diagnosis of vertebral disease. The main idea of the proposed method is a feature augmentation strategy. Firstly, principal components analysis (PCA) is utilized to find out the principal features described pathologies. After that, discrete wavelet transform (DWT) is used to augment saliency features and remove noise data. Through computing the fisher criterion function of linear discriminant analysis (LDA) is to adaptively select the optimal wavelet decomposition level, thereby significantly increasing the inter-class separability. Finally, the support vector machine (SVM) classifier based one-versus-one (OVO) decomposition strategy with nonlinear kernel is employed to classify the augmented features. The proposed method is implemented and compared with other methods. The experimental results show that the proposed method can accurately classify vertebral pathologies and achieve better classification performance.