European Congress of Radiology 2022 - 13 July 2022

Prospective post-marketing surveillance of AI for breast cancer screening in clinical practice

- Danalyn Byng et al.

Purpose:

Prospectively-collected post-market surveillance (PMS) data can provide a multifaceted picture of the robustness of an AI device in clinical practice. We introduce an integrated AI system with live monitoring to support PMS in breast cancer screening.

Methods:

Routinely-documented data on screen-detected cancers and recalls from 6 German screening units using a CE-marked AI system was collected over a 9 month period in 2020/21. Age- and density-stratified crude unadjusted cancer detection and recall rates were calculated and compared between studies read with, and without AI support using Pearson’s chi-squared test. Mixed-effects logistic regression was used to investigate the relationship between screening with AI and likelihood of cancer detection or recall. Models were weighted by inverse propensity scores, adjusted for important risk factors (age, breast density, incident or prevalent screening) and clustered by reader ID.

Results:

Data on N=23,453 studies read with AI, and N=37,019 studies not read with AI were documented. Unadjusted cancer detection rates and recall rates across subgroups were higher for studies read with AI. When controlling for important risk factors, the likelihood of cancer detection was higher, and the likelihood of recall was lower for studies screened with AI vs. studies not screened with AI (odds ratio (OR) for cancer detection 1.11, 95% confidence interval (CI) 0.94–1.31, OR for recall 0.96, 95% CI 0.90–1.03).

Conclusion:

There is a trend towards a higher cancer detection rate and lower recall rate among studies read with AI. Continuous live monitoring in breast screening units, analogous to PMS, can provide invaluable information to safely roll-out AI systems.

Authors:

D. Byng¹, S. Bunk¹, D. Schüler,¹ M. Brehmer^1,2, T. Töllner³, L. Umutlu², K. Pinker-Domenig⁴, C. Leibig¹