Development and Validation of an AI-Assisted Image Acquisition Protocol in Radiography

Thesis Abstract

The study addresses variability in radiographic image quality and patient dose arising from non-standardized image acquisition workflows, proposing an AI-assisted protocol to optimize exposure parameters, positioning guidance, and automatic technique selection within clinical radiography practice. The aim is to design, implement, and validate an AI-driven image acquisition protocol that improves diagnostic image quality while reducing unnecessary radiation exposure. Specific objectives include (1) to develop a supervised machine learning model that recommends exposure factors (kVp, mAs), focal spot size, and automatic positioning cues based on patient anatomy and clinical indication; (2) to integrate the model into a pilot radiography workflow with real-time feedback and ergonomic user interfaces; (3) to evaluate image quality, patient dose, and workflow efficiency before and after implementation; (4) to assess radiographer acceptance and perceived usability; and (5) to validate the protocol across multiple radiology departments with diverse patient populations. A convergent mixed-methods design is employed. The quantitative component utilizes a multicenter quasi-experimental design involving 12 radiography rooms across three tertiary hospitals, with a total sample of 1,200 adult patients indicated for chest, abdomen, and extremity radiographs. The AI model is trained on a retrospective dataset of 80,000 radiographs with associated exposure parameters, body habitus, and clinical indications, and validated on a prospective cohort of 400 radiographs. Image quality is assessed using a standardized 5-point radiologist scoring rubric and objective metrics such as signal-to-noise ratio (SNR) and modulation transfer function (MTF). Patient dose is quantified via dose-area product (DAP) and entrance skin dose (ESD). Statistical analyses include multilevel mixed-effects regression to examine the impact of the protocol on image quality and dose, controlling for patient size, anatomy, and technique history; ANOVA to compare pre- and post-implementation groups; and ROC analysis to evaluate diagnostic adequacy of AI-assisted acquisitions. The qualitative component comprises semi-structured interviews with 24 radiographers and 8 radiology department leads, analyzed via thematic analysis guided by the Technology Acceptance Model (TAM) and the Diffusion of Innovations framework to elucidate usability and adoption factors. Data collection instruments include (a) radiographic acquisition logs embedded in the imaging workflow to capture exposure parameters, positioning cues, and AI recommendations; (b) a validated image quality assessment form completed by three independent radiologists; (c) a dosimetry data extraction tool retrieving DAP and ESD from patient records; (d) structured interview guides for radiographers and department leaders; and (e) a System Usability Scale (SUS) and TAM-based questionnaire to quantify perceived usefulness and ease of use. Validity and reliability are addressed through interrater reliability testing for image quality scores (kappa) and instrument pilot testing prior to full deployment. The analytical framework integrates the AI model’s performance metrics (accuracy, precision, recall) with clinical outcomes (image suitability for interpretation, need for retakes) to determine overall protocol efficacy. Expected findings indicate that the AI-assisted protocol yields a statistically significant improvement in radiographic image quality (mean score increase of 0.8 on a 5-point scale, p<0.01) and a reduction in average DAP by 12–18% without compromising diagnostic confidence, alongside a 15–25% reduction in repeat imaging due to improved initial acquisitions. Multilevel analyses are anticipated to reveal diminished between-physician variability in technique, with faster throughput in busy departments. The qualitative results are expected to reveal high perceived usefulness and acceptable usability (SUS score >70), tempered by considerations regarding workflow integration, accountability for AI-generated recommendations, and need for ongoing model retraining. The study contributes to knowledge by providing a rigorously validated, generalizable AI-assisted image acquisition protocol that links theoretical models of human–machine interaction, radiographic technique standardization, and safety optimization with empirical performance evidence across real-world settings. It advances the literature on AI-augmented radiography by demonstrating feasibility, effectiveness, and user acceptance in improving diagnostic quality while reducing patient radiation exposure. The main conclusion is that a carefully designed AI-assisted protocol, embedded within routine radiography workflows and supported by continuous learning, can reduce practice variability and dose without sacrificing diagnostic capability. Recommendations include establishing ongoing model monitoring, standardized retraining schedules using multicenter data, formal governance for AI recommendations, and expansion of the protocol to additional anatomical regions and pediatric populations, with further studies to assess long-term patient outcomes and economic implications.