Artificial Intelligence user interface preferences in radiology: A scoping review

Abstract

Introduction/Background

Modern forms of Artificial intelligence (AI) have developed in radiology over the past few years. With the current workforce shortages, in both radiology and radiography professions, AI continues to prove its place in supporting clinical radiology processes. The aim of the scoping review was to investigate the existing literature on the topic of preference of use of artificial intelligence interfaces within a radiology context.

Methods

Using a systematic approach, papers were chosen against an inclusion criterion of addressing radiological AI user interface preferences to be included in the scoping review. Arksey O'Malley's and Levac's framework were used to inform the procedural steps for the scoping review. Four databases were searched including MEDLINE Ovid, Scopus, Web of Science and Engineering Village. Reliability was improved through the involvement of three researchers to select the papers against the inclusion criteria.

Results

Six papers were identified to fit the inclusion criteria of radiological AI user interface preferences. These varied methodologically including two observational studies, two simulated user testing studies, a diagnostic accuracy study and a multi-case study. AI user interfaces were evaluated in two studies. Mixed responses were obtained with some alignment in preference for heatmap image overlays and highly detailed user interfaces are linked to higher preference amongst users. Limited literature exists on AI user interfaces and a lack of research evaluating current AI interface preference, either in post or pre-deployment.

Discussion

The mix of methods used within studies indicated that there is not yet a standardised method for assessing AI tool design and preference within radiology, with common use of a System Usability Scale survey tool in conjunction with another method. There was also a varied response when considering the preferred user interface in radiology, though simple, non-complicated designs were suggested to be ideal by participants.

Conclusion

Medical imaging AI user interface research is essential for the acceptability of AI technology into radiology departments. This scoping review identified the current landscape of AI user interface research within a radiology setting. There is a requirement for more radiology AI research focussing on end user or imaging professional involvement and their preferences. There is an explicit need for further research in the field, due to the lack of standardised outcome measures, lack clear findings regarding ideal user interfaces and lack of inclusion of radiographers. The dearth of studies including radiographers and small sample sizes of participants within these studies identifies the mindset shift required for radiology, and AI vendors alike.