Computerized Classification of Benign and Malignant Breast Lesions on DCE-MRI Utilizing Novel Shape Descriptors

Introduction: Dynamic contrast enhanced (DCE)-MRI has recently emerged as an adjunct screening tool to conventional x-ray mammography due to its high detection rate of malignant lesions. However, DCE-MRI is associated with high interobserver variability, with κ ranging from 0.21 to 0.40 [1]. For the specific task of describing lesion morphology (smooth versus spiculated), there is high interobserver (κ=0.29) and intraobserver (κ=0.22) variability [2]. The development of a computerized decision support tool capable of quantifying differences in lesions morphology may aid in reducing observer variability and accurate breast lesion diagnosis on DCE MRI. In this work, we present a computerized classification system to distinguish benign from malignant breast lesions using shape descriptors on DCE MRI. Our classification utilizes a novel Explicit Shape Descriptors (ESDs) to describe differences between the appearance of lesions on DCE-MRI.