AI-assisted Dose Optimization in Digital Radiography Systems

Thesis Abstract

In digital radiography, rising demands for diagnostic image quality must be balanced against patient radiation safety, yet current dose management practices often rely on static protocols that fail to account for patient-specific factors and workflow variability. This study addresses the problem of suboptimal dose optimization in digital radiography systems, proposing an AI-driven framework that personalizes exposure parameters while preserving diagnostic image quality. The objective is to develop, validate, and evaluate a machine learning–enabled dose optimization module integrated with existing radiography platforms to reduce patient effective dose by at least 20% without compromising lesion detectability or clinical workflow efficiency. Specific objectives include (i) modeling the relationship between exposure indices, patient demographics, and anatomical region using supervised learning; (ii) deriving a real-time dose adaptation policy through reinforcement learning to adjust kVp, mA, and exposure time for common radiographic projections; (iii) assessing image quality using objective metrics (contrast-to-noise ratio, Modulation Transfer Function) and observer performance via a double-stimulus, single-interval protocol with radiologist readers; (iv) evaluating system reliability, validity, and integration feasibility within typical hospital information systems and Picture Archiving and Communication Systems (PACS); and (v) conducting a cost-benefit analysis to determine return on investment and safety implications. Methodologically, the study adopts a mixed-methods design conducted in three phases. Phase I (data collection) involves a retrospective cohort of 12,000 anonymized radiographs across chest, abdomen, and extremity studies from three tertiary hospitals, capturing exposure parameters, patient age, weight, BMI, scanner model, and image quality scores. Phase II (model development) employs supervised learning (random forest, gradient boosting, and deep convolutional neural networks) to predict optimal exposure settings, paired with a reinforcement learning agent (Deep Q-Network) to generate dynamic dose recommendations. Phase III (validation and evaluation) uses a prospective, randomized cross-over trial with 600 patients allocated to AI-assisted versus standard protocol arms, measuring effective dose, image quality metrics (CNR, SNR), lesion conspicuity assessments by blinded radiologists, exam turnaround time, and clinician satisfaction. Data collection instruments include calibrated dose meters, standardized imaging phantoms for objective metrics, DICOM-compatible image quality scoring sheets, and a validated observer study protocol. Instrument validity is established through pilot testing with expert radiologists (n=8) and cross-institutional calibration, while reliability is ensured via test–retest procedures and inter-rater agreement analyses (Fleiss’ kappa). Analytical approaches comprise multivariate regression to quantify associations between exposure parameters and image quality outcomes, mixed-effects models to account for clustering by patient and scanner, and Bayesian optimization to tune the AI dose policy. Model performance is compared against conventional protocols using non-inferiority testing for diagnostic accuracy and superiority testing for dose reduction. Feature importance analyses identify key predictors of optimal dose, while sensitivity analyses examine performance under varying noise conditions and patient body habitus. Theoretical framing draws on the Theory of Planned Behavior to understand radiographers’ acceptance of AI-guided dosing and the Information Processing Theory to interpret image quality decision-making under automated guidance. A conceptual framework illustrating the interaction among patient factors, imaging system, and AI-driven dose policy is presented. Expected findings indicate that AI-assisted dose optimization will achieve substantial dose reductions (mean effective dose reduction ~22–28%) without measurable degradation in objective image quality or radiologist diagnostic performance, while also reducing examination time and variability in exposure practices. The study anticipates robust generalizability across device classes with minimal workflow disruption, supported by high user acceptance and acceptable integration latency (<200 ms per exam). The contribution to knowledge includes a validated, transferable AI-based dose optimization framework, empirical evidence on the trade-offs between dose, image quality, and diagnostic accuracy, and a practical model for deploying adaptable exposure policies within clinical radiography workflows and health information systems. The main conclusion posits that patient-specific, AI-guided exposure control can meaningfully reduce radiation risk while maintaining diagnostic integrity, with recommendations for regulatory standardization, continued monitoring of system performance, and phased scale-up across imaging departments. Suggested future work involves expanding to pediatric populations, integrating with automatic anatomy segmentation for projection-specific dosing, and long-term post-implementation surveillance of dose trends and patient outcomes.