Screening participants’ attitudes to the introduction of artificial intelligence in breast screening

A shortage of radiologists is increasingly putting UK breast cancer screening under strain, ¹ and with more breast radiologists retiring than new radiologists being appointed/trained, ² this burden is set to increase. One possible partial solution is the use of artificial intelligence (AI) in breast screening mammogram interpretation to meet this future need. Various groups have demonstrated the potential use of AI in retrospective studies although few have demonstrated its utility, in situ, in large prospective randomised control trials,^3–5 the conventional way of demonstrating clinical and operational utility.

The next step in this technology’s evolution is to determine how best to implement it. Should AI replace all human readers, partially replace them, or operate as a reader assistant/companion? Population screening relies on the test (and its interpretation) being acceptable to those participating. ⁶ However, little is known about the views of the breast cancer screening population on the use of AI in interpreting breast screening mammograms. Data published recently from a Dutch survey of women aged 16–75 do demonstrate overall good acceptance of AI especially when used alongside human screen readers. ⁷

In October 2020, using a standardised paper questionnaire, we sought to obtain NHS Grampian screening participants’ views on the use of AI in interpreting breast screening mammograms, with the aim of designing a prospective study agreeable to the screening population. The questionnaire was designed with help from social scientists and reviewed by University of Aberdeen ethics committee. Its execution was aided by a local charity/patient group and tested for clarity in a similar population to our sample. The complete questionnaire is available in the supplementary material.

After describing the current UK system of dual reader screening followed by arbitration reading if required, we posed four different AI scenarios and asked participants whether they approved or objected; 364 consecutive screening participants were offered the questionnaire, and 187 (51%) returned completed responses. Responses to the four key questions can be seen in Figure 1. We tested the differences in approval/objection for those that expressed a preference (not neutral) using a Chi-squared approach. Participants significantly approved of the introduction of AI for three out of the four scenarios presented. These scenarios involve both AI and human readers to varying degrees. Of these three scenarios, AI as a reader companion (Scenario 4) and AI replacing one human reader (Scenario 1) met with the most approval, while AI as a triage tool (Scenario 3) met with less approval. On average, participants neither approved nor disapproved of the boldest option (Scenario 2), the complete replacement of human readers. In addition to the scenarios, we asked participants their age; perceived knowledge of AI; and if they had a family history of breast cancer. Those with greater self-assessed knowledge of AI were more likely to approve of its introduction. Family history of breast cancer showed no association with AI approval. Age had a weak positive association with approval for Scenario 3, AI as a triage tool.

Figure 1.

The acceptability of different strategies for the introduction of AI into a dual reporting breast screening service.

The gains of each scenario are yet to be quantified; however, it is clear from these results that most of the screening population approves of (or does not object to) the introduction of AI techniques for breast screening. This approval is larger in a subsample who have some self-perceived knowledge or understanding of AI. It is our understanding that these are the first published data demonstrating the effect that perceived understanding of AI has on the likelihood of acceptance of AI within breast cancer screening mammography. However, the level of perceived understanding may be a function of other factors such as education or socioeconomic status. For this reason, the generalisability of our finding may be limited to similar populations. Taken together, the more potentially disruptive the AI scenario (i.e. less human involvement), the lower the level of approval. This is to be expected with the service and the population entering an exciting but uncertain phase, until in-situ knowledge and confidence are gained with AI tools. The more acceptable scenarios may well be stepping stones to a bolder use of AI with potentially more service gains in the future.

Supplemental Material

sj-pdf-1-msc-10.1177_09691413211001405 - Supplemental material for Screening participants’ attitudes to the introduction of artificial intelligence in breast screening

Supplemental material, sj-pdf-1-msc-10.1177_09691413211001405 for Screening participants’ attitudes to the introduction of artificial intelligence in breast screening by Clarisse F de Vries, Brian E Morrissey, Donna Duggan, Roger T Staff and Gerald Lip in Journal of Medical Screening