Radiomics-Machine Learning Analysis for Discrimination of Malign and Benign Breast Lesions on Mammography Images

Abstract

In this study, it is aimed to investigate the analysis radiomics-machine learning on diagnostic performance in differential malign and benign breast lesions using mammography images. In this retrospective study included 101 patients (40 benign and 61 malign). 195 of region of interests (ROIs) were drawn manually by two expert radiologists. Then, using gray level thresholding and morphological operations techniques, each of ROI were segmented on “MATLAB 2020a” program. 126 radiomic features were extracted for each ROI. For eliminating redundant radiomics features, Kruskal Wallis and Relief feature selection methods were used respectively. A total 44 radiomics features were selected after feature selection process. Logistic regression, naive bayes, support vector machine and k-nearest neighbors machine learning algorithms (ML) were used to as classifiers. 10-fold cross validation was applied to measure and evaluate machine learning models. Accuracy, sensitivity and specificity were used as the primary measures of performance of radiomics-machine learning model. Among the machine learning algorithms, support vector machine had the best performance (93.3%, 95.6%, 91.1%). In addition, we found that the feature selection method improved the performance for all ML models. By building the radiomics-ML based analysis with the optimal feature subset, the performance of discrimination of benign and malign lesions showed excellent results which we believe would be useful for clinical practice.