Impact of AI-based image reconstruction on radiographer diagnostic confidence in emergency CT

Thesis Abstract

The integration of AI-based image reconstruction algorithms in emergency computed tomography (CT) presents a timely challenge to radiographers’ diagnostic confidence, with potential implications for workflow efficiency, interpretation accuracy, and patient safety in high-pressure trauma and acute care settings. Despite rapid technical advancements, there is limited empirical understanding of how AI-enhanced reconstruction influences radiographers’ decision-making processes, perceived image quality, and readiness to override conventional protocols in emergency scenarios. This study aims to quantify and elucidate radiographers’ diagnostic confidence when interpreting AI-reconstructed emergency CT images and to identify factors moderating this confidence, including training, prior experience with AI tools, perceived reliability, and contextual workflow pressures. The objectives are (1) to compare radiographers’ diagnostic confidence across AI-reconstructed and standard-dose/filtered back projection CT images in acute head, thorax, and torso cases; (2) to examine associations between confidence levels and objective interpretive performance (sensitivity, specificity) using a reference standard; (3) to explore radiographers’ perceptions of image quality, artefacts, and decision-making thresholds through qualitative inquiry; and (4) to develop a validated framework for implementing AI-based reconstruction in emergency CT that optimizes radiographer confidence and patient outcomes. A mixed-methods design will be employed. The quantitative strand will recruit 120 radiographers from five tertiary hospitals with diversified casemix, who will assess a standardized set of 60 emergency CT cases (30 AI-reconstructed, 30 conventional) presented in randomized order. Each radiographer will rate diagnostic confidence on a 5-point Likert scale and provide interim interpretations, which will be benchmarked against a consensus reference standard established by a panel of blinded neuroradiologists and radiology residents. The objective diagnostic performance will be analyzed using receiver operating characteristic (ROC) curves, area under the curve (AUC), sensitivity, and specificity, with within-subject comparisons conducted via repeated-measures ANOVA. Multivariate linear regression will identify predictors of diagnostic confidence, including prior AI exposure, perceived artefact burden, and workflow interruptions. The qualitative strand will involve semi-structured interviews with a purposive sub-sample of 20 radiographers to elicit in-depth insights into cognitive processes, trust in AI, and decision thresholds when confronted with AI-reconstructed images. Thematic analysis will be guided by the Technology Acceptance Model (TAM) and the Data-fficient Cognitive Load theory to understand acceptance and cognitive demands. Ethical approval will be obtained from institutional review boards, with informed consent secured from all participants. Data collection will ensure de-identification of patient cases, and AI image sets will be validated against established reference standards to prevent bias. Validity and reliability of instruments will be established through pilot testing, inter-rater agreement (Cohen’s kappa) on the reference standard, and internal consistency checks for the confidence scale (Cronbach’s alpha). Data integration will follow a convergent parallel design, with synthesis achieved through joint display analysis to triangulate quantitative outcomes with qualitative themes. Expected findings include higher or non-inferior diagnostic confidence for AI-reconstructed images in certain anatomical regions, contingent on artefact prevalence and radiographer familiarity with AI tools. It is anticipated that confidence will correlate positively with objective diagnostic performance in AI-assisted cases but may exhibit variance in scenarios with technical artefacts or limited AI interpretability. The study will contribute to knowledge by providing empirical evidence on how AI reconstruction affects radiographers’ confidence, decision thresholds, and real-world interpretation, and by offering a contextually grounded framework for training, workflow integration, and governance of AI tools in emergency CT. The implications for practice include informing credentialing requirements, targeted educational interventions to bolster trust and understanding of AI artefacts, and the development of standardized protocols for AI-assisted emergency CT interpretation. The study will guide policy-makers and hospital administrators in optimizing radiology service delivery, minimizing diagnostic uncertainty, and safeguarding patient safety while leveraging AI-based image reconstruction technologies.