AI-Driven Dose Optimization for Digital Radiography Imaging Systems

Thesis Abstract

This study addresses the challenge of balancing diagnostic image quality with patient safety in digital radiography through the development of an AI-driven dose optimization framework that dynamically adapts exposure parameters to individual patient characteristics and exam types. The aim is to minimize unnecessary radiation while preserving clinically acceptable image quality, thereby reducing stochastic and deterministic risks without compromising diagnostic efficacy. Specific objectives include (1) to quantify the relationship between patient demographics, body habitus, and optimal exposure indices across common radiographic projections; (2) to develop and validate a machine learning model that prescribes tube current, voltage, and exposure time to achieve target image quality metrics; (3) to compare dose and image quality outcomes between the AI-driven protocol and conventional fixed or protocol-based approaches; (4) to assess radiographer acceptance, workflow impact, and potential for integration with existing picture archiving and communication systems (PACS) and dose monitoring software; and (5) to evaluate generalizability across equipment brands and clinical settings. The study adopts a mixed-methods design, combining quantitative model development and validation with qualitative usability evaluation. The population comprises adult patients undergoing standard radiographic examinations (chest, abdomen, pelvis, and extremities) at three tertiary hospitals equipped with digital radiography systems. A stratified sample of 1,200 examinations will be collected, with 600 used to train the optimization model and 600 reserved for prospective validation. Instrumentation includes standardized imaging equipment capable of recording exposure indices (EIs, DAP, entrance surface dose), alongside anonymized patient demographic data and clinical indications. A radiographic image quality assessment protocol, based on a validated 5-point Lombard scale and objective metrics such as contrast-to-noise ratio (CNR) and signal-to-noise ratio (SNR), will be employed. The AI model will be developed using supervised learning algorithms (random forests, gradient boosting, and deep learning regression networks) and optimized through cross-validation and grid search. Feature sets will include patient BMI, age, sex, projection, technique factors, and prior image quality metrics. Model outputs will include recommended tube current, kVp, and exposure time tailored to each examination. Data analysis will proceed in two streams. The quantitative stream will involve regression analyses to quantify dose-quality trade-offs, receiver operating characteristic (ROC) analyses to determine diagnostic thresholds, and Bland–Altman plots to assess agreement between AI-suggested and clinician-chosen exposure parameters. Model performance will be evaluated using RMSE, mean absolute error (MAE), and area under the ROC curve for diagnostic adequacy. Sensitivity analyses will test robustness to missing data and equipment variability. The qualitative stream will use thematic analysis of semi-structured interviews with radiographers to evaluate usability, perceived workflow impact, and trust in AI recommendations, with coding grounded in the Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT). The study is expected to show that AI-driven dose optimization maintains diagnostic image quality within clinically acceptable thresholds while achieving a statistically significant reduction in mean entrance skin dose and DAP by at least 15% across examined projections, compared to conventional protocols. Anticipated findings include strong correlations between patient habitus, projection, and optimal exposure settings, and improved consistency of image quality across operators. The research will contribute to knowledge by providing a validated, transportable decision-support model that integrates with PACS and dose-monitoring systems, thereby enabling real-time dose reduction without compromising diagnostic accuracy. Theoretical underpinnings will draw on the ALARA principle and the Technology Acceptance Model to frame dose optimization adoption. The study’s main conclusion is that AI-driven personalized exposure control is feasible and effective across diverse radiography settings, with measurable dose savings and maintained image quality. Recommendations include phased clinical deployment with ongoing performance auditing, integration of the model into radiography workflows, development of standardized calibration procedures across vendors, and continuous model retraining with new data to sustain performance gains. Further research should examine broader modality integration (e.g., mammography, fluoroscopy) and long-term patient outcomes related to reduced radiation exposure.