The Influence of Categorization, Metrology, and Analysis Methods on the Quality and Variability Characterization of Clinical Imaging

Abstract

Computed tomography (CT) imaging is one of the most widely used diagnostic tools in modern medicine. Despite its clinical value, substantial variability exists in CT imaging practices across institutions, scanner models, protocols, and patient populations. This variability complicates efforts to evaluate imaging performance, ensure consistency, and deliver personalized imaging care. At the same time, the increasing availability of large clinical imaging datasets has created new opportunities for quality monitoring, research, and data-driven optimization. However, the ability to extract meaningful insights from these data remains limited by inconsistencies in how imaging studies are categorized, how key parameters are measured, and how imaging data are analyzed. These challenges hinder the ability to answer fundamental clinical questions, interpret variability in imaging practices, and support quality assurance and benchmarking initiatives. This dissertation investigates how categorization, metrology, and analysis methods influence the characterization and interpretation of quality and variability in clinical CT imaging data.The first component of this work focuses on improving the characterization and interpretation of CT imaging practices through enhanced data categorization, measurement, and analysis within radiation dose monitoring systems (RDMSs). RDMSs are widely used in radiology practices to collect and analyze radiation dose information from clinical imaging systems. However, most commercially available RDMS platforms monitor radiation exposure metrics without incorporating image quality information, limiting their ability to support comprehensive imaging optimization. In this work, RDMS functionality was extended to incorporate image quality metrics alongside radiation dose measurements through the design of role-specific visualization interfaces tailored to radiologists, technologists, and medical physicists. The proposed interface was evaluated using a Likert scale assessment framework to measure the effectiveness of fifteen charts designed to address key clinical questions related to imaging performance. The interface received an overall average usability score of 7.8 out of 10.0, with radiologists rating it highest at 8.4, technologists at 7.6, and medical physicists at 7.5. These results demonstrate how radiation dose assessment can be performed in conjunction with image quality evaluation through customizable visualization tools designed to meet the needs of different radiology professionals. The second component of this dissertation addresses metrology challenges associated with patient size characterization in CT imaging. Image quality in CT is strongly influenced by radiation output and patient attenuation; however, patient body habitus varies widely, and multiple metrics have been proposed to quantify patient size. Using a virtual imaging trial framework, six computational human models representing a range of adult body habitus were imaged using a validated CT simulator (DukeSim, Duke University) modeling a representative clinical CT system (SOMATOM Definition Flash, Siemens). Simulated scans were performed across multiple fixed exposure levels for the head, neck, head-neck, head-neck-shoulder, and chest imaging regions with and without noise. Average noise magnitude values were calculated for each slice (σ ̅_slice) by subtracting each image from its noiseless version. Several patient size metrics were evaluated, including effective diameter (D_E) and water-equivalent diameter (D_W) were calculated from each image slice (D_(E_slice ), D_(W_slice )), at center slice (D_(E_center ), D_(W_center )), averaged across slices (D_(E_average ), D_(W_average )), and as the median across slices (D_(E_median ), D_(W_median )). Patient weight and body mass index (BMI) were also used as surrogates of size. Predictive models were developed to relate image noise, tube current, and patient size. Model accuracy was quantified using the mean absolute percent error (MAPE) and relative root mean square error difference (Δ_RMSE). Predictive performance varied substantially across anatomical regions. D_(W_slice ) showed the lowest prediction errors and was therefore used as the reference metric for evaluating other size measures. In the head region, D_(E_slice ) showed the closest agreement to D_(W_slice ), with MAPE of 4.50% for noise and 9.88% for tube current, and corresponding Δ_RMSE values of 1.17 and 1.61. In the neck and head-neck-shoulder regions, most size metrics exhibited substantially higher errors, with some exceeding 100%, except for D_(E_slice ). In the head-neck region, D_(E_slice ) achieved the lowest errors compared to other size metrics. In the chest region, D_(E_slice ), weight, and BMI exhibited the highest percentage errors in both models. These findings demonstrate that the choice of patient size metric significantly influences the accuracy of noise and tube current predictions and should be selected with consideration of anatomical context. The third component examines how measurement choices influence the estimation of radiation risk surrogates. Using the same virtual imaging framework, simulated CT scans were performed across ten exposure levels with both average and strong tube current modulation settings. Organ dose estimates obtained from Monte Carlo simulations were used to calculate effective dose (E) and risk index (RI) metrics. Multiple patient size metrics were evaluated (the same metrics considered above, except the slice-based metrics) for their ability to predict these radiation risk surrogates across head, neck, head-neck, and chest imaging regions. Predictive performance was assessed using exponential regression and normalized root mean square error (nRMSE). The predictive accuracy of size metrics varied substantially depending on anatomical region, risk metric, and exposure conditions. In the head, D_(E_center ) yielded the lowest error for E, while D_(E_average ), D_(E_median ), and D_(W_average ) best predicted RI. BMI showed the poorest performance for both. In the neck, all metrics performed similarly for E, whereas D_(E_median ) and BMI gave the lowest errors for RI. In the head-neck region, no major differences were observed for E, while BMI predicted RI most accurately and center-based metrics performed worst. In the chest, performance was similar across metrics for E, whereas D_W metrics were the most predictive of RI. Stronger tube current modulation settings also increased variability in prediction accuracy. These findings demonstrate that the predictive value of patient size metrics for CT radiation risk depends on anatomical region, exposure level, and tube current modulation schema, emphasizing how size metric selection affects both the accuracy and uncertainty of radiation risk estimates. The final component of this dissertation addresses the structural representation of imaging data through the development of an ontology-based framework for CT data profiling and characterization. The proposed ontology, Operational Ontology for Radiology (OOR), models the radiology imaging workflow from study ordering through image interpretation using explicitly defined entities, relationships, and annotated data properties that capture imaging tasks, acquisition parameters, and workflow context. To illustrate the analytic implications of categorization, CT examinations mapped to a single registry category in the ACR Dose Index Registry (RPID955: CT abdomen pelvis kidney without then with IV contrast) were analyzed. Although these examinations were treated as a single category within the registry, they originated from five institutional protocols with distinct clinical tasks and indications, contrast phase structures, and anatomic coverage. When all examinations were analyzed together, the median total dose length product (DLP) was 1838.9 mGy·cm and could be interpreted as the typical dose for this examination type. However, when the same examinations were regrouped using ontology defined attributes such as imaging task, contrast phases, and scan coverage, median total DLP values ranged from 1301.4 to 2100.7 mGy·cm across task specific groups. These results demonstrate that analytic conclusions about imaging dose and variability depend strongly on how imaging examinations are categorized. The feasibility of implementing OOR in clinical environments was evaluated by assessing the availability of ontology defined data properties within the clinical imaging environment at one comprehensive academic medical center. Among 156 evaluated properties representing key elements of the CT imaging workflow, 48.1% were directly available in structured form, 33.3% required additional standardization, 7.1% were indirectly available but standardized, and 11.5% were not readily available in current systems. These findings demonstrate that a substantial portion of the proposed ontology can be supported using existing clinical data infrastructures while also identifying areas where improved standardization is needed to enable consistent large-scale analysis of imaging practices. Collectively, this work demonstrates that the interpretation of variability in clinical imaging depends critically on how imaging data are categorized, how measurements are defined, and how analyses are performed. By integrating improved monitoring tools, evaluating measurement strategies for patient size characterization, and developing an ontology framework for imaging data representation, this dissertation provides methods to support more transparent and consistent characterization of CT imaging practices. These contributions enable more interpretable benchmarking of imaging performance, support data-driven quality improvement, and establish a framework for scalable analysis of clinical imaging data.