Sources of image texture variation in tomographic breast imaging

Abstract

Texture features have been explored and studied over the last decades providing technical insights that have helped improve a wide variety of fields involving medical and other types of imaging. However there remains a need to examine estimation details, and robustness as significant and new information could be uncovered. Understanding the interactions of imaging system variation and object features in texture formation can provide a corner stone in the advancement of new image processing techniques and acquisition technologies. In this work, we evaluate these questions for digital breast tomosynthesis (DBT) a partial angle tomographic breast imaging system. Recently, our group showed for the first time a correlation between second order texture features and human observer detection performance in digital images. We also showed that second order texture features commonly used as “radiomic” metrics can change with multiple acquisition and reconstruction methods. Here we focus on issues related to robustness in estimating these features. Specifically, we aim to understand how Haralick’s GLCM texture features, used in radiomic models as predictors, change under different estimation conditions in simulated DBT images. We attempt to understand and analyze the effects that different breast densities, pixel distance offsets, ROI window sizes and filtering have on GLCM texture features calculations.