About the generalizability of deep learning based image quality assessment in mammography

Abstract

One method of assessing the image quality of a mammography unit is to estimate a contrast-detail-curve (CDC) that is obtained from images of a technical phantom. It has been proposed to estimate this CDC by using an end-to-end neural network (NN) which only needs one image to determine the CDC. That approach, however, has been developed on the basis of images of one single mammography unit. In this work, we train NNs on synthetic images of contrast-detail phantoms for mammography and test the so-trained NNs on images that are obtained from real mammography units. The goal of this paper is to demonstrate that such a deep learning approach is capable to generalize to predict CDCs for various real mammography units. Our experiments cover various manufacturers and the proposed approach is shown to work across different NN architectures and preprocessing methods which highlights its generalizability.