篇名 | An Image Classification Method for Digital Breast Tomosynthesis Based on Combined Texture Feature Extraction |
---|---|
卷期 | 31:4 |
作者 | You-Ming Wang 、 Jia-Qi Miao 、 Han-Mei Zhang |
頁次 | 001-013 |
關鍵字 | digital breast tomosynthesis image 、 feature selection 、 gray level co-occurrence matrix 、 support vector machine 、 EI 、 MEDLINE 、 Scopus |
出刊日期 | 202008 |
DOI | 10.3966/199115992020083104001 |
A new combined algorithm is proposed for the texture feature extraction of digital breast tomosynthesis (DBT) image based on gray level co-occurrence matrix (GLCM), Tamura and Relief algorithms. The disadvantage of the GLCM algorithm is that the global information including the pixel dependencies among the textures can not be fully utilized. The presentation of Tamura algorithm can overcome the shortcomings of the GLCM algorithm on the way of integrating the characteristics of human visual information into the process of feature extraction. In this paper, the texture features are screened by Relief algorithm, which is set as a criterion to select the optimal features. The DBT image is preprocessed to reduce the noise and increase the contrast by bilateral filtering, contrast-limited adaptive histogram equalization and L0 gradient filtering. The moving window is used to traverse the image. In each moving window, the energy, contrast, correlation, entropy, inverse difference moment (IDM) and correlation of the GLCM algorithm and the features of the Tamura algorithm including the coarseness, contrast and directivity are extracted, respectively. Each set of Tamura texture features is integrated into GLCM texture features to construct a fused texture feature space and the fused texture features are filtered for the optimal features by Relief algorithm. The filtered texture features are input into the support vector machine (SVM) for the image classification. Compared with other extraction algorithms, the proposed algorithm can improve the accuracy of feature extraction and recognition. The results show that the classification accuracy and efficiency of the proposed algorithm are much improved compared with SVM, GLCM, Tamura and GLCM-Tamura algorithms.