Development of an AI-powered Radiographic Image Quality Assessment System

Thesis Abstract

The quality of radiographic images significantly influences diagnostic accuracy and patient outcomes, yet current assessment practices are largely subjective, inconsistent, and labor-intensive, highlighting a critical need for automated, reliable, and scalable quality evaluation systems in radiography. This study aims to develop an artificial intelligence (AI)-powered radiographic image quality assessment system that leverages deep learning algorithms to objectively evaluate image quality parameters such as resolution, contrast, noise, and artifacts. The specific objectives include designing a robust AI model employing convolutional neural networks (CNNs), validating its performance against expert radiologist assessments, and determining the system’s potential integration into routine radiology workflows to enhance diagnostic accuracy and operational efficiency. The research adopts a cross-sectional, experimental design incorporating both qualitative and quantitative methods. The population comprises 300 digital radiographic images obtained from a tertiary hospital equipped with state-of-the-art imaging equipment, along with 15 seasoned radiologists serving as evaluators of image quality. A stratified random sampling technique is employed to select images across different anatomical regions and imaging modalities to ensure diverse data representation. Data collection involves gathering the images from the hospital’s Picture Archiving and Communication System (PACS) and obtaining quality ratings from radiologists using a standardized scoring rubric. Supplementary data are collected through structured interviews with radiology department stakeholders to contextualize system integration challenges. The AI model is developed using a training set of 200 images, with the remaining 100 images reserved for validation and testing. The system's performance is assessed through quantitative metrics such as accuracy, sensitivity, specificity, precision, and the area under the receiver operating characteristic (ROC) curve. Machine learning techniques like transfer learning with pre-trained CNN architectures (e.g., ResNet50 and InceptionV3) are applied to extract features from the images. Comparative analysis involves regression techniques and Bland-Altman plots to evaluate agreement between AI assessments and expert ratings. The study further employs thematic analysis to interpret stakeholder insights regarding system adoption. This comprehensive approach ensures rigorous evaluation of the AI model’s diagnostic utility and practical applicability. Expected findings indicate that the AI-based system can accurately classify radiographic images based on their quality, achieving an accuracy exceeding 90% and an ROC curve area above 0.95, thus demonstrating reliability comparable to expert radiologists. The system is anticipated to significantly reduce assessment time by automating the evaluation process and minimize inter-observer variability, thereby standardizing quality assurance in radiography. Furthermore, the integration of this system is expected to facilitate immediate feedback to radiographers, enabling real-time quality control and reducing repeat scans. This research contributes to the existing body of knowledge by providing a validated, scalable AI framework for objective radiographic image quality assessment, filling a notable gap in automated quality control literature. It advances theoretical understanding by contextualizing AI application within the socio-technical environment of radiology, supported by the Theory of Technological Adoption and the Socio-Technical Systems Theory. The study also offers practical insights into implementing AI-driven solutions in clinical settings, including considerations for system robustness, user trust, and workflow integration. The main conclusion underscores the potential of AI to enhance the precision, consistency, and efficiency of radiographic image quality assessment, ultimately improving diagnostic outcomes. Recommendations include further multi-center validation studies, development of user-friendly interfaces for clinical adoption, and ongoing training for radiology staff to maximize system benefits. Future research could explore integrating AI-based quality assessment with advanced diagnostic AI systems to develop comprehensive radiology decision-support platforms.