The survey had 87 nationally accredited respondents; 73 (83.9%) completed the survey in full. Most (61%) had over 10 years’ experience, and 77% were consultant radiologists. Nineteen participants provided comments. While just over a third (37%) described their understanding of AI as good or excellent, 63% had a positive or strongly positive view of AI use in screening. One respondent indicated: “I am in favour of adopting AI in mammogram reporting.” Another respondent stated: “AI has a role in breast screening and would help to alert. [AI would] [a]lso help with personnel shortage.” Most (82%) had not been previously involved in procuring similar medical software for their organisation.
Figure 1 shows participants’ responses to which AI implementation scenario they would prefer. Respondents preferred partial replacement (AI replaces one human reader) over other AI implementation scenarios. They objected to the total replacement scenario, while views on the triage and companion scenarios were mixed.
Two respondents suggested alternative AI implementation scenarios. One comment stated that “[i]t would be great to have AI tested against previous interval cancers as this is one of the few things that will influence outcomes / breast cancer mortality in the screened population” and that AI could be used “on all those cases given normal results by the readers as a safety net system prior to results being sent out.” The second response suggested that double reading with AI would not save a lot of radiology time, and that AI would be better used to maximise image quality, decide whether to perform breast imaging with tomosynthesis, pre-read symptomatic mammograms, and focus on risk and masking from breast density/parenchyma.
Approximately half of the respondents thought first readers (52%) and second readers (51%) should have access to the AI opinion. Most respondents (68%) thought that third readers or an arbitration panel should have access to the AI opinion.
Figure 2 shows participants’ responses to what evidence they think would support AI introduction into their workplace. Most respondents rated national guidelines (77%), studies using a nationally representative dataset (65%) and independent prospective studies (60%) as essential to support the introduction of AI into clinical practice. Vendor generated evidence, however, was considered to have limited value. Most participants indicated that evidence generated from local data was either essential (43%) or desirable (42%).
Seven comments discussed the need for additional evidence and validation of AI breast screening tools, including different software, the threshold for recall and readers’ interactions with the AI. Related comments stated: “Replies non-committal because I want to see the evidence first!”, “I am strongly in favour of adopting AI in screening mammography reading once it has been validated and made user friendly” and “AI has so far shown excellent results with better than human sensitivity and specificity but needs input of robust data and validation tests locally and nationally.” One respondent suggested that a national working group of AI specialists and screen readers should be developed through the Royal College of Radiologists to evaluate and test the various AI systems and ways of using them on large datasets. They added: “National guidelines are vital to ensure it is used in the optimal manner and to provide medicolegal protection.”
The view that the second specialist is blinded to the first reader’s opinion was held by 45% of participants; 54% indicated that it was their view that the specialist should be blinded to the AI opinion. Two respondents indicated that they were unsure whether the question on the blinding of the second specialist to the first reader referred to whether they are currently blinded or whether they should be blinded.
Figure 3 shows participants’ responses to how they would rank the given AI representation options. Respondents preferred a region of suspicion superimposed on the image over other shown AI representation options.
Readers with a greater self-reported knowledge of AI were more likely to view the use of AI in medical screening as positive (ρ = 0.496, p < 0.001). Self-reported knowledge of AI was not significantly associated with approval of any of the AI implementation scenarios (p > 0.05).
The remaining free-text comments predominantly related to concerns regarding the introduction of AI into breast screening, including lack of planning for the needed infrastructure, and potential negative effects on screen readers, patients, and screening centres. One participant indicated that it is “[i]mportant that training of future mammographic readers is not forgotten, as AI cannot assess patients.” Relatedly, one respondent stated: “There needs to be widespread understanding of the limitations of AI as I am afraid that readers will have too much faith in its abilities.” Another participant commented: “AI will decrease specificity and increase recall rates. Radiologists will be left to cope with the fall out at assessment clinics. How can centres be assessed for QA [quality assurance] if AI is introduced?” One respondent indicated that AI is “[d]ifficult to introduce” and “buy-in from most radiologist[s]” must be obtained before introducing AI in breast screening nationally. They further stated that ethical questions should be answered in a FAQ (frequently asked questions document) to reassure screen readers. One screen reader responded: “I believe it is inevitable that AI will be introduced over the next few years and we need to ensure it is done so in the most effective manner for the breast screening programme.”