Evaluation of a digital breast tomosynthesis cancer detection AI algorithm using the personal performance in mammographic screening scheme (PERFORMS)

Purpose: With the growing uptake of Digital Breast Tomosynthesis (DBT) for screening and the development of DBT Artificial Intelligence (AI), we aim to compare the performance of a DBT AI model as a standalone reader to that of a large cohort of breast screening readers from the PROSPECTS Trial, using the PERFORMS external quality assurance (EQA) scheme. 

Materials and methods: 75 combined DBT and Synthetic 2D mammography (S2D) screening cases were collated into a PERFORMS test-set. The test-set was completed by 88 breast screening readers from 7 NHS hospitals participating in the PROSPECTS trial. The same set was analysed by Hologic Genius AI® Detection 2.0 software. Standalone AI performance was benchmarked against the performance of the reader cohort, and differences between AI and human scores were assessed using the Wilcoxon signed rank test (α = 0.05). 

Results: The reader cohort had a median of 12 years (IQR: 4–17) experience in breast screening and 5 years (IQR: 2–7) experience using DBT in screening. Human readers achieved a median Area Under the Receiver Operating Characteristic Curve (AUC) of 0.934, median sensitivity of 92.1% and specificity of 88.4%. In comparison, the AI model achieved an AUC of 0.935 (P = 0.13), and a sensitivity of 97.4% (p < 0.001) and specificity of 71.4% (p < 0.001) at the preset manufacturer threshold. 

Conclusion: The study showed that the overall standalone performance of the DBT AI model in terms of AUC was not significantly different than that of a large cohort of specialist breast screening readers in the UK.

(British society of breast radiology annual scientific meeting abstracts - abstract O3. BSBR Annual Scientific Meeting 2025 Brighton, UK 9–11 November 2025 https://www.delegate-reg.co.uk/bsbr2025.)