Radiomics analysis on contrast-enhanced spectral mammography images for breast cancer diagnosis: A pilot study

Abstract

Contrast-enhanced spectral mammography is one of the latest diagnostic tool for breast care; therefore, the literature is poor in radiomics image analysis useful to drive the development of automatic diagnostic support systems. In this work, we propose a preliminary exploratory analysis to evaluate the impact of different sets of textural features in the discrimination of benign andmalignant breast lesions. The analysis is performed on 55 ROIs extracted from 51 patients referred to Istituto Tumori "Giovanni Paolo II" of Bari (Italy) fromthe breast cancer screening phase betweenMarch 2017 and June 2018. We extracted feature sets by calculating statisticalmeasures on original ROIs, gradiented images,Haar decompositions of the same original ROIs, and on gray-level co-occurrencematrices of the each sub-ROI obtained byHaar transform. First, we evaluated the overall impact of each feature set on the diagnosis through a principal component analysis by training a support vectormachine classifier. Then, in order to identify a sub-set for each set of featureswith higher diagnostic power,we developed a feature importance analysis bymeans of wrapper and embeddedmethods. Finally,we trained an SVMclassifier on each sub-set of previously selected features to compare their classification performances with respect to those of the overall set. We found a sub-set of significant features extracted fromthe original ROIs with a diagnostic accuracy greater than 80%. The features extracted fromeach sub-ROI decomposed by two levels of Haar transformwere predictive only when theywere all usedwithout any selection, reaching the bestmean accuracy of about 80%. Moreover, most of the significant features calculated by HAAR decompositions and their GLCMs were extracted from recombined CESMimages. Our pilot study suggested that textural features could provide complementary information about the characterization of breast lesions. In particular, we found a sub-set of significant features extracted fromthe original ROIs, gradiented ROI images, and GLCMs calculated fromeach sub-ROI previously decomposed by theHaar transform.