Building Blocks for Machine Learning in Medical Imaging

Our prior research has shown a correlation between several second order image texture features and human observer studies in tomographic breast images. Digital breast tomosynthesis (DBT) acquisition parameters were tested for both sensitivity and specificity performance for low contrast mass detection via localization receiver operating characteristic (LROC) studies. These acquisition parameters dictate the optimal imaging dose vs. signal detectability for an imaging geometry or reconstruction algorithm being tested. Simulated phantom images were generated using different acquisition parameters as part of the virtual clinical trial platform being developed in our laboratory. These correlations between the image texture features and human attention can aid in efficient system designs capable of image classification and localization of signals such as calcification clusters and low contrast cancer. In this study, we explore image features linked to texture parameters in order to characterize the human observer’s regions of interest (ROIs). Our findings have the potential to unveil essential features and building blocks for a fast and efficient machine learning algorithm in tomographic breast imaging.