Purpose: Model observers have traditionally been applied to simple tasks where the lesion location is known exactly (LKE) or the lesion morphology is known exactly (SKE). However, these conditions are far from a clinical scenario. Here, we propose an observer model to detect masses of different size. This model can be used in virtual clinical trials comparing Digital Mammography (DM) and Digital Breast Tomosynthesis (DBT) images.
Methods: We generated irregular spiculated masses of seven different sizes. Using the VICTRE pipeline, we generated 2270 DM and DBT samples for each size. Average glandular dose was consistent with the clinical dose levels and with the dose reported in the VICTRE project. We trained a Channelized Hotelling Observer (CHO) template per size by isolating the mass from the rest of the tissues. With target-only images, we trained multiple CHO models for different channel sizes and computed the performance using full sample sizes containing all the breast tissues in target-present and target-absent images. Finally, we analyzed the AUC for both modalities at different lesion sizes for 30 reader models.
Results: The best channel set combination for DBT and DM show a consistent increase in detectability with DBT always higher than DM. Both modalities seem to track in terms of the benefits of having larger signals in the AUC values. As lesion size increases, the size of the channels should increase accordingly. We compared the AUC for a wide range of channel sizes and found that DBT images are more sensitive to channel sizes than DM.
Conclusion: As the size of breast lesions increases, we found a monotonic increase in AUC for both modalities. The DBT modality shows the best AUC performance for all sizes for the best corresponding channel combination.