Classifying breast cancer regions in microscopic image using texture analysis and neural network

This study proposes and evaluates a neural network (NN) classifier for dividing the histological structures (HS) in breast cancer (BC) microscopic image into two region types: cancer or normal. Cancer region included positive cells and negative cells while normal region included stromal cells and lymphocyte. The classification task using a back propagation learning algorithm is applied to the multilayer perceptron architecture of NN classifiers. To yield a high classification performance, the main focus of interests is feature extraction task using four texture features: correlation, autocorrelation, the information measure of correlation and fractal dimension. A combination of these texture features is used in 60 images for training data set and 104 images for testing data set. The comparison of performances between each texture feature and the combination of them has been reported. The results show that the best classification accuracy obtained from the all features is 94.23%. This indicated that the texture analysis and NN classifier are feasible for dividing the HS in BC microscopic images and can be applied to improve and to develop an accurate cell counting of computer-aided systems for BC diagnosis.