Classification of Breast Masses via Transformation of Features using Kernel Principal Component Analysis

T. Mu, A.K. Nandi (UK), and R.M. Rangayyan (Canada)


Breast masses, pattern classification, feature transforma tion, kernel principal component analysis, Fisher’s linear discriminant analysis.


Several studies on classification of breast masses in mam mograms have shown that shape features are highly suc cessful in discriminating between malignant breast tumors and benign masses, as compared to edge-sharpness and tex ture features. However, the extraction of shape features re quires accurate contours which are not easy to obtain auto matically. In this paper, we propose to apply kernel princi pal component analysis (KPCA) to the problem of classi fication of breast masses, aiming to improve the discrimi nating power of each single feature in an expanded feature space derived from a centered kernel matrix. We also aim to improve the discriminating capability of different fea ture combinations in other transformed, more informative, lower-dimensional feature spaces, especially with the edge sharpness and texture features. Fisher’s linear discrimi nant analysis (FLDA) is employed to evaluate the classi fication capability of the transformed features via KPCA. The methods were tested with a set of 57 regions in mam mograms, of which 20 are related to malignant tumors and 37 to benign masses, represented using one shape feature, three edge-sharpness features, and 14 texture features. The classification performance of the edge-sharpness and tex ture features, via KPCA transformation, was significantly improved from 0.75 to 0.85 in terms of the area under the receiver operating characteristics curve.

Important Links:

Go Back