Mammographic density estimation through permutation entropy

Abstract

The American College of Radiology, through its committee on BI-RADS (Breast Imaging Study Data and Reporting System), has concluded that breast density is more clinically important as an indicator of concealment of possible breast lesions than as a quantifier of cancer risk, due to the lack of robust descriptors for detecting diverse types of density. In this work, new descriptors for mammographic density estimation based on the Permutation Entropy (PE) algorithm are developed and assessed. PE is a measure of complexity initially proposed for chaotic time series, particularly in the presence of dynamic and observational noise. We propose different novel algorithms to adapt the concept of PE from time series to images, to characterize the level of roughness. Once the characteristic vector for each mammogram was obtained, we trained a multilayer feedforward neural network as a classifier, to evaluate the potentiality of the set of descriptors as mammographic density characterizers, in accordance with the BI-RADS nomenclature. The results show that these descriptors have remarkable success rates in the classification of densities and especially they generalize with good coincidence percentages for cases of extreme densities. The categorization of extremely dense breasts is of special interest because of their clinical importance to assign more intensive monitoring or more complex studies to the patients who present it.