Novel feature extraction based on DCT-DOST features for classification of Digital Breast Tomosynthesis images into benign and malignant tumors

Abstract

Background: Breast cancer is caused by the uncontrolled growth of abnormal cells, resulting in a mass in the breast tissue. Early detection of the disease can significantly improve its prognosis and treatment outcome. Digital Breast Tomosynthesis, a three-dimensional imaging technology of the breast tissue, is emerging as a standard for breast imaging with improved screening and diagnostic results. The additional information obtained from tomosynthesis reduces the misleading effects of tissue overlap and improves the detection, identification, and localization of abnormalities. Benign-malignant breast tumors classification at early detection is a crucial step to improve diagnosis and prolong patient survival. The aim of this study is to classify benign-malignancy of masses in Digital Breast Tomosynthesis images based on new DCT-DOST and Radiomic features on an open database of TCIA that consist of 224 lesion bounding boxes. Methods: In order to effectively obtain both DCT-DOST and Radiomics characteristics, a 2D central slice of the DBT image, which encompasses a substantial anatomical size of the breast tumor, was employed. In the pre-processing stage, rescale intensity method was used to improve contrast and image quality. After that, by using a binary mask, the mass of the breast tissue was subjected to segmentation. Then, using radiomic and DCT-DOST features, new features were extracted from the images. In addition, we also investigated the importance of feature selection and class balancing. By pooling the radiomics and DCT-DOST features into a hybrid feature set, we investigated the compatibility of these two sets with respect to benign malignancy prediction. Results: Finally, for classification using Random Forest algorithm, K-Nearest Neighbor and Support Vector Machine, it was shown that the best result of the evaluation metrics are respectively equal to 87.80%, 78.51%, 82.78% and 75.19% in terms of mean AUC, Accuracy, Sensitivity and Specificity which was obtained by Random Forest classifier. Conclusion: The findings from the empirical analysis reveal that integrating DCT-DOST and Radiomic features into the same learning algorithm improves the discrimination power.