Decision Support System in Identification of Lesions in the Dense Breast on Digital Mammograms

Abstract

Background The most common cancer affecting women globally is breast cancer. The most effective and extensively used tool for identifying breast abnormalities in the early stage is mammography screening. However, it is less effective when there is high breast density. Therefore, the radiologist considers a sonomammography in addition to mammography to identify and characterize the lesion. Therefore, we aim at modelling a decision support system to identify a lesion in a dense breast using a mammogram and thereby avoiding additional sonomammogram. Methods In this work, the image pixels were pre-processed to produce super-pixels that are very meaningful and easy to analyse. Texture features extracted using a rotation invariant local binary pattern (RI-LBP) approach from super-pixels. Feature Selection algorithm was used to extract the top 5 features. The performance of various machine learning models was studied, and the best model was used for the prediction of the presence or absence of lesions in the mammogram images. Results The findings suggest that breast masses localized in the dense background can be described by features produced from RI-LBP super-pixel patterns in very effective and efficient ways. The experiment showed that with 500 super-pixels and only 5 RI-LBP features, the Support Vector machine model with gaussian kernel yielded an accuracy value of 90.38% and an area under the receiver operating characteristic (ROC) curve score of 0.96 for the mammogram dataset.