Comparison of model-observer and human-observer performance for breast tomosynthesis: Effect of reconstruction and acquisition parameters

摘要

The problem of optimizing the acquisition and reconstruction parameters for breast-cancer detection with digital breast tomosynthesis (DBT) is becoming increasingly important due to the potential of DBT for clinical screening. Ideally, one wants a set of parameters suitable for both microcalcification (MC) and mass detection that specifies the lowest possible radiation dose to the patient. Attacking this multiparametric optimization problem using human-observer studies (which are the gold standard) would be very expensive. On the other hand, there are numerous limitations to having existing mathematical model observers as replacements. Our aim is to develop a model observer that can reliably mimic human observers at clinically realistic DBT detection tasks. In this paper, we present a novel visual-search (VS) model observer for MC detection and localization. Validation of this observer against human data was carried out in a study with simulated DBT test images. Radiation dose was a study parameter, with tested acquisition levels of 0.7, 1.0 and 1.5 mGy. All test images were reconstructed with a penalized-maximum-likelihood reconstruction method. Good agreement at all three dose levels was obtained between the VS and human observers. We believe that this new model observer has the potential to take the field of image-quality research in a new direction with a number of practical clinical ramifications.