An intelligent early breast cancer detaction using statistical feature generation

Breast cancer is one of the main causes of women death worldwide. Breast tumor is an early stage of cancer that locates in cells of a human breast. Early breast cancer detection greatly increases the chances for early diagnosis. Towards this, it is very crucial to propose an intelligent classifier to detect breast cancer in the early stage. This thesis proposes a preliminary research to detect the size and location of the breast cancer in the early stage. UWB signals are transmitted by the antenna from one side of breast phantom and received from other side, controls by PC. These UWB time domain signals are converted into frequency domain signals using Fast Fourier Transform. Both time and frequency domain signals are converted into digital values. For size classification, data is analysed by using the proposed statistical feature generator method. Initially, Principal Component Analysis is performed on the data and then, is normalized into ten different data normalization methods. The data dimension is reduced by reducing the principal components. Out of ten data normalization method, only five are chosen statistically. Ten features, combination of time (linear and non-linear) and frequency (linear) features are extracted from each of the five normalized dataset. Out of 50 extracted features, ten features are selected based on the statistical test. These features are fused together using feature fusion techniques. 10-HybridFeature, 8-HybridFeature and 3-HybridFeature datasets are developed using the proposed features and then are tested with three different supervised classifiers for size classification. Among these three datasets, 8-HybridFeature dataset performs better for all three classifiers. Experiment shows Support Vector Machine_RBF (94.07%) classifier performs better compared to Naïve Bayes (91.98%) and Probabilistic Neural Network (81.64%) classifiers by using 8-HybridFeature dataset in terms of accuracy. The proposed method performs better than previous work by improving 11.8% of accuracy. For location classification in terms of x, y, and z coordinates, the data is normalized using binary normalization and classified using unsupervised classifier (k-means clustering). The k-means clustering classifier exhibits average accuracy of 80.49%. Based on the classification result, the best feature and best classifier are selected to proposed Breast Cancer Detection (BCD) algorithm. The proposed BCD algorithm consists of two parts. The first part is for size detection using supervised classifier and second part is for location detection using unsupervised classifier. Finally, the detected size and location are visualized in 2D and 3D environment.