A Novel Classifier Design Algorithm MatMHKS And Its Application On Breast Cancer Detection

Classifier is an important ingredient in pattern recognition. Among all the classifiers, linear classifiers are paid great attention in statistical pattern recognition due to their simplicity and easy expansibility to nonlinear classifiers. Until now, ther e are several linear classifiers, e.g., Perceptron, Relaxation, MSE and Ho-Kashyap(HK) algorithm. HK is not robust to outliers. The modified HK with Square approximation of the misclassification errors (MHKS) tries to avoid this shortcoming and adopts similar principle to the support vector machine to maximize the separation margin. The operating object of all these linear classifiers is vector pattern, i.e., before applying them, any non-vector pattern should be firstly vectorized into a vector pattern. But, such a vectorization will bring at least three potential problems: 1) Structural or local contextual infor mation may be broken down; 2) The higher the dimension of input pattern, the more me mory space are needed for the weight vector related to a classifier; 3) When the dimension of a vector pattern is very high and while the sample size is small, it is easy to be overtrained. In this paper, inspired by the method of feature extraction directly based on matrix patterns and the advantage of MHKS, we develop a new MHKS classifier based on matrix patterns (MatMHKS). The method can mitigate the above shortcomings.We also make a further try of applying the algorithm proposed above to breast cancer detection. In medical detection, the cost of false positive (FP) and false negative (FN) is different, regarding this point, we introduce a cost coefficient to MatMHKS, and design a new algorithm called CS-MatMHKS. In early detection of breast cancer, digital ma mmography is considered to be the most reliable method, the presence of microcalcification clusters(MCCs) is an important sign for the early detection. In this thesis, we first try to extract some useful features of MCCs recommended by experts, and then perform classification directly by CS-MatMHKS, consequently, more information can be saved and the rate between FP and FN can be controlled and traded-off.