TU-CD-207-06: Effect of Reconstruction Methods On the Evaluation of Digital Breast Tomosynthesis (DBT) Acquisition Parameters Using Human and Model Observers




We previously showed that the choice of reconstruction methods did not affect the optimization for DBT acquisition parameters (angular span (ASP) and number of views (NOV)). We considered the task of detecting spherical lesions using a channelized Hoteling observer (CHO). The goal of this work is to validate our model observer results using human reader studies.


Sets of DBT projections of multiple lesion-present and lesion-absent digital breast phantoms were generated using a virtual DBT system at various combinations of ASP and NOV. Each set of projections was reconstructed into DBT slices with several reconstruction methods. In this study, we investigated the performance trends for optimizing ASP at 9 views and optimizing NOV at 20o span with FBP and SART reconstructions. 140 sub-images for each geometry were extracted from the simulated DBT slices for human reading. Human observers scored each sub-image according to the confidence level for presence of a lesion at a known location. The area under the ROC curve (AUC) was estimated as a function of ASP or NOV for both reconstruction methods. AUC curves for the CHOs using Laguerre-Gauss (LG) and Square channels were also plotted for comparison.


Our initial results based on one reader indicate that the performance trends as a function of acquisition parameters are similar between the outcomes from FBP and SART reconstructions. Both the LG and the Square CHOs predict the reader's performance trends well. Results for a total of 4 readers will be presented at the time of conference.


The preliminary results validate that the choice of reconstruction methods does not impact the optimization of DBT acquisition parameters. If confirmed, our results indicate that a CHO with LG or Square channels can predict human performance trends for detecting a spherical lesion in breast-mimicking tomosynthesis backgrounds.