Scale-Space DCE-MRI Radiomics Analysis Based on Gabor Filters for Predicting Breast Cancer Therapy Response

Abstract

Radiomics-based studies have created an unprecedented momentum in computational medical imaging over the last years by significantly advancing and empowering correlational and predictive quantitative studies in numerous clinical applications. An important element of this exciting field of research especially in oncology is multi-scale texture analysis since it can effectively describe tissue heterogeneity, which is highly informative for clinical diagnosis and prognosis. There are however, several concerns regarding the plethora of radiomics features used in the literature especially regarding their performance consistency across studies. Since many studies use software packages that yield multi-scale texture features it makes sense to investigate the scale-space performance of texture candidate biomarkers under the hypothesis that significant texture markers may have a more persistent scale-space performance. To this end, this study proposes a methodology for the extraction of Gabor multi-scale and orientation texture DCE-MRI radiomics for predicting breast cancer complete response to neoadjuvant therapy. More specifically, a Gabor filter bank was created using four different orientations and ten different scales and then firstorder and second-order texture features were extracted for each scale-orientation data representation. The performance of all these features was evaluated under a generalized repeated cross-validation framework in a scale-space fashion using extreme gradient boosting classifiers.