The Effect of Test-Set Training on Mammography Screen-Reading Performance

Abstract

Aims: This thesis aims to validate the impact of test-set training on screening mammography readers' clinical performance. It initially reviews the literature on the role of test sets and clinical audits in maintaining reader performance and presents three studies that utilise clinical audit data for evaluation.

Methods: Data from BreastScreen NSW between 2008 and 2018 were analysed, focusing on participants in the BREAST test-set programme. The first study compared pre- and post-training annual audit metrics for 24 radiologists. The second study used a control group of peers without test-set training for comparison. The third study employed statistical tests to analyse the correlation between test-set metrics and changes in cancer detection rates for 41 participants.

Results: Overall, radiologists showed significant improvements in recall rates, positive predictive value (PPV), and specificity post-training. Higher-volume readers also improved cancer detection rates. Compared to non-test-set peers, BREAST participants exhibited additional significant increases in overall and invasive cancer detection rates. Multiple correlations were found between test-set performance and changes in cancer detection rates, with a significant regression equation established based on test-set sensitivity and specificity.

Conclusion: Experience improves reader performance, but those participating in test sets show more significant improvements, particularly in cancer detection rates. Sensitivity and specificity were identified as the key test-set performance indicators for predicting improvements.