Impact of AI-based Dose Optimization on Patient Radiation Exposure in Clinical Radiography

Thesis Abstract

The study addresses the persistent challenge of unnecessary patient radiation exposure in clinical radiography by evaluating the effectiveness of artificial intelligence (AI)-based dose optimization strategies integrated into radiographic workflows. Despite advances in imaging technologies, variability in exposure settings and protocol adherence contribute to suboptimal dose management, potentially compromising patient safety and diagnostic quality. The aim is to quantify reductions in effective dose and assess diagnostic image quality when AI-driven dose optimization is implemented in routine radiography. Specific objectives include (1) measuring changes in patient entrance surface dose (ESD) and effective dose (E) before and after AI integration, (2) evaluating image quality and diagnostic acceptability using standardized scoring protocols, (3) determining the influence of AI recommendations on examiner adherence to dose optimization guidelines, (4) identifying operational facilitators and barriers to implementation in diverse clinical settings, and (5) exploring radiographers’ cognitive workload and perceived trust in AI-driven tools. A mixed-methods, multi-site study will be conducted in three tertiary care hospitals with diverse patient populations. The quantitative component will employ a quasi-experimental, pre-post design comparing two six-month periods baseline (n ? 3,000 radiographic examinations) and post-implementation (n ? 3,200 examinations). AI-based dose optimization will be deployed within existing image acquisition pipelines, with algorithms recommending automatic exposure parameter adjustments while preserving diagnostic requirements. Data sources will include exposure parameters (kV, mA, exposure time), automatic exposure control (AEC) responses, patient demographics, and DICOM metadata. ESD will be recorded from dosimeters and clinic records; effective dose will be estimated using established conversion coefficients. Image quality will be assessed by two blinded radiologists using the European Guidelines on Minimum Standards for Radiographic Imaging, with a subset of 600 images re-evaluated for inter-rater reliability (Cohen’s kappa). The qualitative component will use semi-structured interviews and focus groups with 20 radiographers and 10 radiology department supervisors to explore implementation experiences, perceived AI trust, and workflow integration. The theoretical framework will draw on the Technology Acceptance Model (TAM) and the Safety-II perspective to interpret how AI supports safe dose practices without compromising diagnostic value. Quantitative analyses will include paired t-tests and repeated-measures ANOVA to compare ESD and effective dose across periods, multivariate linear regression to adjust for patient age, body mass index, and clinical indication, and ROC analysis to examine diagnostic quality thresholds. Inter-rater reliability for image quality will be evaluated with intraclass correlation coefficients (ICCs). The qualitative data will be analyzed using thematic analysis guided by Braun and Clarke’s approach, with coding conducted independently by two researchers and triangulated with quantitative findings. A conceptual model will be developed to illustrate causal pathways among AI optimization, dose metrics, image quality, and user acceptance, drawing on the Kano model for feature prioritization. Key expected findings include statistically significant reductions in mean ESD and effective dose (targeting at least a 15% reduction post-implementation) without meaningful degradation in diagnostic quality (image quality scores remaining within established acceptability ranges). It is anticipated that AI guidance will increase protocol adherence rates by radiographers, reduce exposure variability, and lower cognitive workload when integrated into intuitive interfaces. The study will contribute to knowledge by providing empirical evidence on the real-world effectiveness of AI-driven dose optimization, clarifying the relationship between automated parameter recommendations and clinical outcomes, and informing guidelines for safe deployment in radiography. The main conclusion is that AI-based dose optimization can meaningfully reduce patient radiation exposure while maintaining diagnostic integrity, provided that careful integration, ongoing monitoring, and clinician trust-building are in place. Recommendations include standardizing AI training datasets to ensure generalizability, implementing continuous quality assurance programs with dose audits, incorporating user-centric interface designs, and conducting longitudinal studies to assess long-term safety, cost-effectiveness, and impact on workflow efficiency.