Systemic Integration of Retrospective Analysis for Optimizing CT Imaging Practice

Abstract

Computed tomography (CT) plays a central role in modern medical imaging, where maintaining appropriate image quality while controlling radiation dose is essential for accurate diagnosis and patient safety. Current clinical monitoring systems primarily focus on radiation dose metrics such as CT dose index (CTDIvol) and dose-length product (DLP). However, these metrics do not directly reflect diagnostic image quality and may fail to detect clinically meaningful variations in image quality across scanners, protocols, and patient populations. The overall objective of this dissertation is to advance CT image quality monitoring by introducing and evaluating methods that move beyond dose-centered metrics toward task-based and organ-specific assessments.The first study investigates the limitations of dose-centered monitoring by examining detectability-based image quality metrics in a large clinical dataset. A task-based detectability index was estimated for clinical CT series, and its relationship with image acquisition and patient factors was evaluated using a multivariable model rather than radiation dose alone. By incorporating multiple predictors known to affect detectability, the model provided a more comprehensive estimate of expected image quality under varying imaging conditions. Prediction intervals derived from this model were used to define the expected range of detectability and to identify series with detectability values outside this range as outliers. The findings demonstrate that dose metrics alone are insufficient for identifying clinically meaningful image quality outliers and support the use of multivariable, task-based monitoring for improved CT quality assessment. The second study focuses on the development and optimization of an organ-specific image noise metric, termed the Organ Noise Index (ONI). Based on Global Noise Index (GNI) framework, ONI quantifies local noise characteristics within segmented organs and requires several key computational parameters, including Hounsfield unit thresholds, kernel size, and histogram bin width. Using clinical CT data, these parameters were systematically evaluated to determine configurations that produce stable and reliable noise estimates across different organs. The third study applies the optimized ONI methodology to a large clinical dataset consisting of 3133 CT image series. Organ-specific noise measurements were obtained for several abdominal and thoracic organs and compared with global noise metrics and patient size indicators. The analysis showed strong agreement between organ noise and global noise for relatively homogeneous soft tissues such as the liver, spleen, kidneys, and urinary bladder, while the lungs exhibited distinct behavior due to their heterogeneous anatomical structure. These findings demonstrate the feasibility of large-scale organ-specific noise analysis and highlight differences in noise behavior across organs and imaging conditions. In summary, this dissertation demonstrates that CT image quality monitoring can be improved by incorporating task-based detectability metrics and organ-specific noise assessment. The methods developed and evaluated in this work provide a framework for more clinically relevant evaluation of CT image quality and may contribute to improved protocol optimization and quality assurance in clinical CT imaging.