Perceptually Relevant Measurements in Tomosynthesis Imaging

Tissue structures and imaging system parameters are major factors that influence the ability of radiologists to find and identify malignancies. For example, parameters relevant to mammography like radiation dose, breast density and thickness, pixel size, arc span and number of projections in the case of digital breast tomosynthesis (aka 3D mammography) can alter the cancer screening diagnosis. Therefore, it is necessary to find which configuration is best so that least number of malignancies are missed by the radiologists. They examine hundreds of radiographs every day and sometimes there is a potential to misclassify the images. By understanding the image properties in relation to human perceptual mechanisms, we can find the optimal configurations to reduce misdiagnosis. The focus of this thesis is to characterize digital breast tomosynthesis (DBT) images using image quality metrics and eye gaze analysis and to explore how these metrics relate to abnormality detection performance and diagnostic errors. The first part of this thesis evaluates the relevance of noise power spectrum (NPS) based metrics in characterizing the DBT images and their relation to human observer diagnosis performance. The NPS based power-law exponent beta was hypothesized to quantify tissue structural overlap and to use as a surrogate for diagnostic performance in x-ray imaging. We showed that random noise and system configurations influence this parameter and do not correlate with human observer performance in DBT imaging and hence cannot be used as a surrogate for diagnostic performance while the power-law magnitude (K) showed a strong correlation. Next, we investigated the impact of phantom structural variations on the estimation of DBT optimal configurations in virtual imaging trials (VITs). Our results indicate that phantoms should be designed to resemble the patients’ anatomical structures accurately and should be evaluated for their realism and sufficiency in the use of VITs. Finally, to further our understanding of the relation between different system parameters and human perceptual mechanisms, gaze analysis was conducted. To this end, we developed a graphical user interface for an eye-tracking tool to conduct eye-tracking studies and estimate eye gaze metrics and fixation regions. Our results indicate that along with diagnostic performance other perceptual aspects such as image reading time can be considered for system optimization. Our gaze analysis suggests that gaze metrics correlate with diagnostic performance and task difficulty and could help understand the difficulty levels in a VIT.