This recommendation has been updated during the latest Guidelines Development Group meeting, considering available evidence until July 2023. 

Should double reading with support by artificial intelligence vs. double reading without support by artificial intelligence be used to read mammograms using digital mammography (2DFFDM) or digital breast tomosynthesis for early detection of breast cancer in mammography screening programmes?

The majority of the GDG agreed that the balance of desirable and undesirable effects favours the use of double reading with AI support. The overall certainty of the evidence is very low because of uncertainty of the accuracy of the readings, and the absence of data for the downstream impact on breast cancer and mortality. None of the studies evaluated the use of the readings in a screening programme or followed individuals over time.

The evidence shows that when using double reading supported by AI, there may be slightly more false positives than double reading without support. However, there may be one more breast cancer detected for every 1 000 women screened when double reading is supported by AI, which outweighed the false positives. 

Although the addition of AI as a support to double reading is probably feasible and acceptable, it may increase costs, and therefore its adoption should be conditional on sustainability. The GDG also noted that, compared to previous/older computer-aided detection systems, the accuracy of AI has improved over the last few years and is expected to improve. The GDG will carefully monitor the evolution of this field and consider this question again in the future.

The majority of the GDG agreed with the recommendation: 6 members voted for a conditional recommendation for double reading with support by artificial intelligence, 2 members voted for a conditional recommendation for either double reading with support by artificial intelligence or double reading without support by artificial intelligence, and 2 members abstained.

Before acquiring and implementing AI technology, an evaluation of the legal context for its use is needed. In addition, thorough quality control of the AI system should be conducted and might include the assessment of data provided by the vendor (e.g. testing and validation results, ROC curves, and number of images utilised for training the algorithm).

Quality assurance and quality improvement of AI technology should be conducted and might require storing of AI and consensus results.

• Prospectively conducted studies are needed which also measure patient important outcomes, such as breast cancer and mortality. Information about how and when to use AI for most benefit is needed and should be evaluated in studies.
• Linkage with cancer registries could assist in connecting the long-term follow-up data to the intervention.
• Further research, including retrospective analyses of existing data, should support identifying which subgroups (e.g. lobular cancers, DCIS) artificial intelligence performs better.