Machine learning based computer aided diagnosis system for classification of breast masses in mammograms

Abstract

Breast cancer continues to be the most common cancer in the fastest developing and the developed nations. Early detection by using mammography has been proven as the best prognosis. Computer Aided Diagnosis (CAD) systems are being used as second reader for the analysis and interpretation of mammogram images. In the last two decades, although breast cancer incidence has increased by many folds but unfortunately the progress in this field has almost stagnated. Therefore, the CAD systems need to be improved to be considered useful. In this study, a machine learning based CAD system for segmentation and classification of breast masses have been proposed. The IRMA Version of DDSM dataset has been used for experimentation and evaluation of the proposed system. Exact breast masses were segmented from manually extracted ROIs of 700∗700 pixels by employing an improved seeded region growing algorithm. Various geometry and texture features were computed from the segmented mass lesions and corresponding ROIs respectively. The classification performances of nine state-of-the-art classifiers namely K-Nearest Neighbour (KNN), Support Vector Machine (SVM), Gaussian Mixture Model (GMM), Multi-class Support Vector Machine (mSVM), Decision Tree (DT), Discriminate Analysis (DA), Naive Bayes (NB), Random Forest (RF), Ensemble Tree (ET) have been investigated in this study. On evaluating the experimental results for all the classifiers, highest classification accuracy is obtained with SVM classifier. The experimental results reveal that the proposed improved seeded region growing approach has been proven helpful in improving the classification performance of the proposed system.