Artificial intelligence in radiographic image interpretation

: an investigation of the impact of an AI model on diagnostic accuracy, trust and decision switching in radiographers

Abstract

This thesis investigated the impact of artificial intelligence on radiographers’ image interpretation. A literature review revealed heterogeneity in the way that studies investigating AI for fracture identification were reported. Impressive performances were reported in all studies (n=30) however, the metrics used in many papers may be unfamiliar to clinicians, such as prevalence agnostic metrics to allow for imbalanced datasets.

A survey was carried out to establish UK radiographers’ perceptions of clinical AI (n=411) and found that UK radiographers do not feel well prepared for the future with AI technologies in the clinical radiology setting and desire more training. Less than a third of Reporting Radiographers responding indicated that they were confident in explaining the AI output to service users and other healthcare practitioners. Indication of overall performance of the system and visual explainability of the AI focus would serve to increase users’ trust.

The impact of AI feedback on student and qualified radiographers’ rates of decision-switching and automation bias was investigated and found that students were more likely than radiographers to change their mind following AI feedback. Heatmap feedback caused increased rates of decision switching across both groups. Both groups followed incorrect advice (automation bias). This was more prevalent in the student group.

The final study investigated factors impacting reporting radiographers’ trust when using AI. Participants agreed with binary diagnosis more often than the heatmap feedback (86.7% agreement with AI diagnosis) and disagreed with the heatmap feedback on 45.8% of pathological cases (n=66). Correlations were found between trust and agreement with both binary and heatmap AI feedback.

This work will assist with development and procurement, encouraging consideration of forms of AI feedback which will be beneficial to users. The findings should be used to guide educators on the contents of programmes to upskill the workforce and educate students in AI.

Date of Award	Nov 2023
Original language	English
Supervisor	Raymond Bond (Supervisor), Ciara Hughes (Supervisor), Sonyia Mc Fadden (Supervisor) & Jonathan McConnell (Supervisor)