Reporting radiographers’ interaction with Artificial Intelligence—How do different forms of AI feedback impact trust and decision switching?

Abstract

Artificial Intelligence (AI) has been increasingly integrated into healthcare settings, including the radiology department to aid radiographic image interpretation, including reporting by radiographers. Trust has been cited as a barrier to effective clinical implementation of AI. Appropriating trust will be important in the future with AI to ensure the ethical use of these systems for the benefit of the patient, clinician and health services. Means of explainable AI, such as heatmaps have been proposed to increase AI transparency and trust by elucidating which parts of image the AI ‘focussed on’ when making its decision. The aim of this novel study was to quantify the impact of different forms of AI feedback on the expert clinicians’ trust. Whilst this study was conducted in the UK, it has potential international application and impact for AI interface design, either globally or in countries with similar cultural and/or economic status to the UK. A convolutional neural network was built for this study; trained, validated and tested on a publicly available dataset of MUsculoskeletal RAdiographs (MURA), with binary diagnoses and Gradient Class Activation Maps (GradCAM) as outputs. Reporting radiographers (n = 12) were recruited to this study from all four regions of the UK. Qualtrics was used to present each participant with a total of 18 complete examinations from the MURA test dataset (each examination contained more than one radiographic image). Participants were presented with the images first, images with heatmaps next and finally an AI binary diagnosis in a sequential order. Perception of trust in the AI systems was obtained following the presentation of each heatmap and binary feedback. The participants were asked to indicate whether they would change their mind (or decision switch) in response to the AI feedback. Participants disagreed with the AI heatmaps for the abnormal examinations 45.8% of the time and agreed with binary feedback on 86.7% of examinations (26/30 presentations).’Only two participants indicated that they would decision switch in response to all AI feedback (GradCAM and binary) (0.7%, n = 2) across all datasets. 22.2% (n = 32) of participants agreed with the localisation of pathology on the heatmap. The level of agreement with the GradCAM and binary diagnosis was found to be correlated with trust (GradCAM:—.515;—.584, significant large negative correlation at 0.01 level (p = < .01 and—.309;—.369, significant medium negative correlation at .01 level (p = < .01) for GradCAM and binary diagnosis respectively). This study shows that the extent of agreement with both AI binary diagnosis and heatmap is correlated with trust in AI for the participants in this study, where greater agreement with the form of AI feedback is associated with greater trust in AI, in particular in the heatmap form of AI feedback. Forms of explainable AI should be developed with cognisance of the need for precision and accuracy in localisation to promote appropriate trust in clinical end users.