AI-Driven Dose Optimization for Pediatric Radiography Imaging Protocols

Thesis Abstract

In pediatric radiography, excessive or poorly optimized radiation exposure poses heightened cancer risk and deterministic effects, while suboptimal image quality can necessitate repeat examinations, undermining diagnostic accuracy and patient safety. This study develops and validates an AI-driven dose optimization framework designed to tailor imaging protocols to individual pediatric patients, balancing dose minimization with diagnostic adequacy. The aim is to reduce effective dose by at least 20% on average without compromising radiographic quality or clinical workflow efficiency. Specific objectives include (1) to identify key patient- and protocol-related factors influencing dose and image quality, (2) to develop a machine learning model that predicts optimal exposure parameters (kVp, mAs, and automatic exposure control settings) conditioned on age, weight, height, body habitus, and clinical indication, (3) to integrate the model within a decision-support module linked to existing radiography workstations, and (4) to evaluate the framework across multiple common pediatric radiography procedures (chest, abdomen, and extremities) in terms of dose metrics, image quality scores, referral turnaround time, and technologist workload. A mixed-methods, multi-center study design is employed. The population comprises pediatric patients aged 0–18 years undergoing standard radiographic examinations at five tertiary care centers. A stratified sample of 1,200 procedures (240 per center; balanced across age bands neonates, infants, toddlers, school-age, and adolescents) will be collected prospectively over 12 months. Data collection instruments include (i) standardized radiographic exposure logs, automatic dose monitors, and DICOM metadata; (ii) image quality assessment using a structured five-point Likert scale rated by three blinded radiologists per study; (iii) technologist workflow metrics captured via workstation analytics; and (iv) a clinician-specified diagnostic confidence questionnaire. The methodology integrates an AI modeling phase and a prospective validation phase. The AI component will employ supervised learning (gradient boosting and deep neural networks) to predict optimal exposure settings from patient morphology and clinical indication, incorporating domain-inspired constraints derived from the ALARA principle and the Reference Image Quality Criteria. The model will be trained on 70% of the retrospective dataset and validated on the remaining 30%, with nested cross-validation to optimize hyperparameters. Feature engineering will include age- and weight-normalized indices, body mass index proxies, and anatomical region encoding. A Bayesian uncertainty quantification layer will accompany predictions to inform decision-making under clinical risk. The integration phase will implement a decision-support module with explainable AI (SHAP values) to present rationale behind parameter recommendations and safety thresholds. Data analysis will combine quantitative and qualitative approaches. Dose reductions will be analyzed using paired t-tests and mixed-effects linear models to account for clustering by center and modality, with effect sizes reported as Cohen’s d. Image quality and diagnostic confidence will be assessed using nonparametric ANOVA (Kruskal-Wallis) and intraclass correlation coefficients to evaluate inter-rater reliability. A cost-benefit analysis will estimate potential reductions in repeat imaging and throughput changes. Sensitivity analyses will test robustness to varying thresholds for acceptable image quality. Theoretical framing will be anchored in the ALARA principle, Information Processing Theory for decision-support usability, and the Technology Acceptance Model to gauge clinician adoption. Expected findings include statistically significant reductions in mean organ-dose-equivalent metrics (e.g., effective dose per exam) across procedures, without degradation in image quality or diagnostic confidence, and with marginal improvements in workflow efficiency. The study aims to demonstrate that AI-driven dose optimization provides reliable, interpretable guidance to technologists while preserving radiographic standards. The contribution to knowledge lies in the development of a scalable, clinically validated framework for personalized radiographic exposure in pediatrics, integrating radiologic physics, machine learning, and human–AI collaboration. The study will inform policy on pediatric imaging protocols and contribute to guideline updates for dose optimization. Practical recommendations will cover implementation steps, required training for radiology staff, and considerations for regulatory compliance. The main conclusion expected is that AI-assisted, patient-specific protocol optimization can achieve meaningful dose reductions without compromising diagnostic utility, and future work will explore real-time adaptive imaging and broader procedural applicability.