A National Audit of Mammography Systems Settings That May Affect the Output of Artificial Intelligence Software

Abstract

Background: Artificial Intelligence (AI) software in mammography is trained on a set of processed images and may be less effective when applied to images acquired on different systems or systems with different processing and/or acquisition settings. The aim of this work was to undertake a retrospective audit of a large number of mammography systems in the United Kingdom and identify the number of differences in image acquisition and processing factors. Methods: Images of the TORMAM phantom are acquired as part of the routine quality control programme. Data from the DICOM header of these images were extracted to provide a snapshot in time of the system configurations. A longitudinal audit of DICOM header data for all of the Hologic systems was tested by one medical physics department (MPD1) over 14 years. Results: We received results from 28 UK medical physics services for 498 systems. There were 7 different models of mammography systems, each with up to 7 different versions of acquisition workstation software. Each mammographic model had multiple image processing versions, including bespoke settings. The GE had two dose settings, while Siemens systems had a range of doses from 80% to 150% of the standard dose. In the longitudinal audit, there were between 2 and 6 software versions in concurrent use on the Hologic systems tested by MPD1. Conclusions: This study showed the heterogeneity of system setup across the UK in a single year, as well as changes to system setup over time. These differences may affect the outcomes of both AI and human readers. There are responsibilities on AI suppliers, mammography equipment manufacturers, breast-screening units, and medical physics services to ensure outcomes are not adversely affected by differences or changes in mammography equipment configurations.