Smart Radiography Diagnostics: AI-Powered Image Quality Optimization System

Thesis Abstract

The study addresses the critical challenge of variability in radiographic image quality due to disparate acquisition protocols, patient factors, and operator-dependent decisions, which collectively impact diagnostic accuracy and radiation safety. It proposes an AI-powered image quality optimization system (IQOS) integrated into standard radiography workflows to automate quality assessment, guide exposure parameter selection, and provide real-time feedback to technologists. The aim is to enhance image quality consistency while minimizing patient dose. Specific objectives are (1) to develop a multimodal AI model that predicts optimal exposure settings and positioning guidance from initial scout images and patient metadata; (2) to implement a real-time quality score and corrective action module that flags suboptimal images and suggests corrective steps; (3) to evaluate the system’s impact on diagnostic confidence, image repeat rates, and patient dose across diverse clinical settings; and (4) to assess user acceptance and integration challenges among radiographers. The study adopts a mixed-methods research design comprising a quantitative phase to train and validate the AI model and a qualitative phase to examine user experience. The population includes radiography departments across five tertiary hospitals, with a total of 1,200 consecutive chest and extremity radiographs used for model training and 600 subsequent examinations for prospective evaluation. Data collection instruments consist of (a) a structured dataset of anonymized radiographic images with corresponding exposure parameters, image quality annotations by expert radiologists, and patient demographics; (b) a computerized workflow log capturing machine settings, exposure indices, and image acquisition times; (c) standardized surveys and semi-structured interviews to gauge radiographer acceptance, perceived usefulness, and perceived ease of use; and (d) a validation panel assessment of image quality pre- and post- IQOS intervention. The primary analysis employs deep learning architectures, including a convolutional neural network (CNN) for image quality prediction and a reinforcement learning component to optimize exposure parameter recommendations. Regression analysis and paired t-tests will compare presystem and postsystem metrics such as mean dose-area product (DAP), image repeat rate, and diagnostic confidence scores. A multivariate ANOVA will assess differences across hospital sites and imaging planes. The qualitative data will be analyzed using thematic analysis to extract patterns related to workflow integration, trust in AI recommendations, and training needs. The study will also apply the Technology Acceptance Model (TAM) to interpret user acceptance outcomes. Expected findings indicate that IQOS reduces image repetition by 18–25%, lowers average patient dose by 12–20% without compromising diagnostic quality, and increases radiographers’ confidence in exposure decisions through transparent AI rationale and actionable guidance. The AI model is anticipated to generalize across scanner brands and imaging protocols with transfer learning and domain adaptation techniques, though site-specific calibration may be required. The study contributes to knowledge by bridging AI-driven image quality optimization with practical radiography workflows, offering a replicable framework for safe, dose-conscious imaging through automated parameter guidance and real-time quality feedback. It advances theoretical understanding of technology-mediated workflow optimization in medical imaging by integrating constructivist notions of professional practice with contemporary AI governance and human-in-the-loop design. The final conclusion is that AI-powered IQOS can substantially enhance consistency in radiographic quality and dose management when embedded within coherent training, robust validation, and clinician-centered interfaces. Recommendations include establishing standardized validation datasets across institutions, implementing ongoing model monitoring for drift, providing structured radiographer training on AI outputs, and developing governance policies to address accountability, data privacy, and explainability.