Anatomical noise and visual realism assessment in simulated mammography and digital breast tomosynthesis images from data‐driven breast phantoms

Background Virtual imaging trials in medicine (VITMs) rely on in silico platforms that simulate tests on volunteers. In VITMs, a simulation software program replicates the imaging apparatus, and the patient is replaced by a digital phantom. Recent innovations in the field of x‐ray breast imaging supported a widespread development of VITM platforms and the need for appropriate evaluations and validations. In this context, digital patients are usually created via mathematical models based on a priori anatomical assumptions, that permit the reproduction of a large population variability with arbitrarily fine spatial resolutions. On the other hand, data‐driven digital phantoms—derived from 3D images of the breast—feature realistic shape and glandular tissue distribution; however, they are limited by the scarce availability of 3D breast images with relatively high image quality and by the spatial resolution of the used clinical data. These are largely adopted in dosimetry evaluations, but no single generally adopted criterion for realism evaluations—which depend on the intended clinical task—in simulated images exists. Purpose This work aims to evaluate the realism of digital mammography (DM) and digital breast tomosynthesis (DBT) images generated for VITMs by adopting data‐driven digital breast phantoms via objective parameters and subjective analysis. Methods DM and DBT images were simulated via an open‐source simulation platform. The cohort of data‐driven digital breast phantoms included 126 cases from an open access database. The subjective analysis was conducted via a two‐alternative forced choice (2AFC) test and a Visual Grading Analysis (VGA) involving four radiologists. In the 2AFC test, radiologists were asked to recognize regions of interest (ROIs) extracted from simulated images, by comparison with a second set from clinical images. VGA evaluated image realisms based on a scoring scale. The objective analysis was conducted by evaluating the β parameter in images, this characterizing the magnitude of the anatomical noise. Results The β parameter was 2.71 ± 0.40 in DM and 3.09 ± 0.28 in DBT. From the analysis of the noise spectra of DM images, the presence of artifacts caused by the phantom voxelization was evident and prominent for thinner breasts. Radiologists distinguished ROIs from the simulated dataset from those from a clinical dataset in the vast majority of cases. Low pass filters on simulated images had little influence on DBT images; however, the sensitivity in detecting simulated images was no lower than 83%. VGA average scoring acknowledged some degree of image realism, outlining the presence of artifacts or unrealistic image regions. Conclusions Objective evaluation reported that the β parameter evaluated for DM and DBT images simulated by adopting data‐driven digital breast phantoms was in line with that from clinical images. A 2AFC test demonstrated that radiologists distinguish between simulated and real images. The subjective analysis also permitted the identification of the main features characterizing simulated images that include voxelization artifacts, the absence of fine structures and of a realistic adipose pattern in simulated images. However, realism should be verified with regard to a specific question of interest under investigation in a VITM.