Systematical Analysis and Pathological Classification of Breast Cancer from Mammographic Images with Using Specific Machine Learning Methods

Abstract

For years, breast cancer has been a serious problem and malignant tumor case primarily causes death of women all around the world. In this paper, a computer based breast tumor analysis and pathological case classification system has been achieved and some novelties are included to the image processing methods, especially in segmentation and base frequency distribution acquisition of the processed image and classification part. First, the possible noises and artifacts are eliminated by using common filtering. Second, the filtered images are segmented with integrating gray level Image Processing methods. Then, these images (ROIs) are converted to the base frequency distribution images with using Fast Fourier Transform (FFT) and Lab&HSV color spaces. The most important key for these images is frequency distribution can be obtained with specific color tones and totally 100 images (50 benign-50 malignant) are accumulated to fed the two different Machine Learning models in literature such as Probabilistic Neural Network as Learning Vector Quantization (LVQ) and Support Vector Regression (SVR) for classification of Benign and Malignant cases without the need for additional medical data. Then the performance of the proposed system is analyzed with 30 different test images (15 benign-15 malignant) according to the metrics like accuracy, sensitivity, specificity, precision, F-score and area under the ROC curve (AUC score). The experimental results on the open access mammogram image set show that discriminating between Benign and Malignant cases can be achieved with an important success rate as 91.38% with LVQ and %.92 with SVR.