In 2011, 2012 and 2013, 14 radiologists independently assessed a year-specific BREAST mammographic test-set. Radiologists read a different single test-set once each year, with each comprising 60 digital mammogram cases. Radiologists marked the location of suspected lesions without computer-aided diagnosis (CAD) and assigned a confidence rating of 2 for benign and 3–5 for malignant lesions. The mean sensitivity, specificity, location sensitivity, JAFROC FOM and ROC AUC were calculated. A Kruskal–Wallis test was used to compare the readings for the 14 radiologists across the 3 years. Wilcoxon signed rank test was used to assess comparison between pairs of years. Relationships between changes in performance and radiologist characteristics were examined using a Spearman's test.

Results

Significant increases were noted in mean sensitivity (P = 0.01), specificity (P = 0.01), location sensitivity (P = 0.001) and JAFROC FOM (P = 0.001) between 2011 and 2012. Between 2012 and 2013, significant improvements were noted in mean sensitivity (P = 0.003), specificity (P = 0.002), location sensitivity (P = 0.02), JAFROC FOM (P = 0.005) and ROC AUC (P = 0.008). No statistically significant correlations were shown between the levels of improvement and radiologists' characteristics.

Conclusion

Radiologists' who undertake the BREAST programme demonstrate significant improvements in test-set performance during a 3-year period, highlighting the value of ongoing education through the use of test-set.