Impact of AI-based Dose Optimization in Radiography at St. Mary's Hospital Radiology Department

Thesis Abstract

This study investigates the impact of AI-based dose optimization on radiographic imaging practices within St. Mary’s Hospital Radiology Department, addressing the persistent challenge of balancing diagnostic image quality with patient radiation exposure in a high-volume clinical setting. The problem centers on the variability of dose recommendations across conventional protocols and the potential of AI-driven algorithms to standardize exposure while preserving diagnostic accuracy, thereby reducing cumulative patient dose and optimizing workflow efficiency. The aim is to evaluate (1) changes in patient- and procedure-specific radiation dose metrics, (2) the effect on image quality and diagnostic confidence, (3) technologist workflow efficiency, and (4) stakeholder acceptance and perceived usability of AI-based dose optimization tools. Specific objectives include quantifying dose metrics (CTDIvol, DLP, and ionizing radiation dose per radiograph) before and after AI implementation; assessing image quality using standardized metrics and blinded radiologist scoring; examining changes in repeat-rate and examination turnaround times; analyzing technologist satisfaction and perceived usability through validated surveys; and exploring barriers to adoption via semi-structured interviews under a theoretical lens of Technology Acceptance Model (TAM) and Diffusion of Innovations (DOI). The methodology adopts a mixed-methods, longitudinal, quasi-experimental design conducted over 18 months, comprising a 6-month pre-implementation baseline and a 12-month post-implementation period. The population includes all radiographic procedures performed in the department, with a purposive sample of 1,200 adult radiographs across chest, abdomen, and extremity studies for quantitative dose and image-quality analysis, and a convenience sample of 40 radiographers for workflow and usability assessments. Data collection instruments include automated dose records from the Radiology Information System (RIS) and Picture Archiving and Communication System (PACS); a standardized image-quality assessment protocol using a 5-point Likert scale validated by prior studies; a technologist workflow survey measuring turnaround times and satisfaction; and semi-structured interview guides aligned with TAM and DOI constructs. Instrument validity is ensured through expert panel review and pilot testing with 5 radiographs per modality, and reliability is established via Cronbach’s alpha for survey scales and inter-rater agreement (Cohen’s kappa) for image-quality scoring. Data analysis employs a two-pronged approach. Quantitative analysis uses descriptive statistics to summarize dose metrics and image-quality scores, paired-sample t-tests or Wilcoxon signed-rank tests for pre- vs post-implementation comparisons, and multivariate linear regression to adjust for confounders such as patient thickness, technique factors, and modality. Repeated-measures ANOVA will examine trajectory effects over time. Mediation analysis will explore whether improvements in diagnostic confidence mediate the relationship between dose optimization and clinical outcomes. Qualitative data from interviews will be analyzed thematically using a priori codes informed by TAM and DOI, with triangulation against quantitative findings to enhance validity. The analysis will be conducted with R and NVivo software, ensuring transparency via preregistered analytic plan and data management procedures. Key expected findings include a statistically significant reduction in mean dose per examination (anticipated 15–25% decline in CTDIvol and DLP for CT-related studies; 10–20% for radiographs), without compromising image quality or diagnostic accuracy, and a reduction in repeat rates by 8–15%. Technologist workflow efficiency is anticipated to improve, evidenced by shorter procedure times and higher satisfaction scores. The study also expects to identify key determinants of AI adoption, with TAM demonstrating perceived usefulness and perceived ease of use as primary predictors of intention to use, moderated by organizational support per DOI constructs. The study contributes to knowledge by providing empirical, context-specific evidence on AI-driven dose optimization in a real-world radiology department, offering a rigorous evaluation framework for dose reduction without sacrificing diagnostic integrity, and detailing facilitators and barriers to adoption in a high-demand clinical environment. Practical recommendations include refining automated exposure protocols, targeted training modules for radiographers, ongoing monitoring of dose metrics, and a structured change-management plan to sustain AI-enabled improvements beyond the study period. The findings are expected to inform policy development for dose optimization, guide vendor decision-making for AI radiography solutions, and contribute to best-practice guidelines for radiology departments pursuing responsible, evidence-based dose management.