Evaluation of image quality and diagnostic accuracy in breast tomosynthesis. Comparison with digital mammography and estimates of radiation risk

Abstract: Recent advances in the digital detector technology have paved the way for modalities such as two-dimensional digital mammography (DM) and three-dimensional breast tomosynthesis (BT). In this work, experimental studies were conducted to investigate the influence of reducing the average glandular dose (AGD) on detection of simulated lesions in DM and to compare the diagnostic accuracy of BT and DM. In addition; an alternative model for fitting ROC curves was validated with the use of clinical data. The influence of reduced AGD (increase in system noise) on detection of simulated lesions was investigated in search-based observer performance studies. Two lesion types were investigated, masses and microcalcifications, at three dose levels: the currently used AGD level at Skåne University Hospital Malmö (1.3 mGy for a standard breast, 100%), 50% and 30% of that. Both masses and microcalcifications were affected by dose reduction as indicated by the consistent decrease in the figure of merits, but microcalcifications were particularly sensitive (p < 0.05). The potential benefit in terms of lives spared because of an AGD reduction was estimated based on a diminished radiation-induced breast cancer risk. Depending on risk model used, three to nine lives would be spared by using half of the AGD per 100,000 women undergoing screening mammography annually from 40 to 49 years of age. The effect of combining the BT view with the corresponding ipsilateral DM view was compared to single-view BT and standard dual-view DM regarding breast cancer detection. The combined modality of BT and DM, was superior to standard dual-view DM (p < 0.05). Single-view BT versus standard dual-view DM was investigated further using a larger sample-size. It was found that the radiologists performed significantly better on single-view BT compared to dual-view DM; average per reader there was 10.4 more breast cancers detected on BT of 95 breast cancers in the study population. The sensitivity of BT was higher than that of DM (average ~ 90% vs. average ~ 79%; 95% confidence interval (CI) of difference: [0.036, 0.108]), while the average false positive fraction was not significantly different (95% CI of difference: [-0.117, 0.010]). An alternative method for fitting ROC curves (search-model, SM) for diagnostic accuracy estimations was compared with those predicted by the widely used proper ROC model (PROPROC) when applied to clinical data. The results were indicative of a strong linear correlation between the SM AUC and PROPROC AUC (R2 = 0.95), shows that a model based on visual search can be used to fit ROC curves.