Predicting Task-specific Performance for Iterative Reconstruction in Computed Tomography
The cross-sectional images of computed tomography (CT) are calculated from a series of projections using reconstruction methods. Recently introduced on clinical CT scanners, iterative reconstruction (IR) method enables potential patient dose reduction with significantly reduced image noise, but is limited by its "waxy" texture and nonlinear nature. To balance the advantages and disadvantages of IR, evaluations are needed with diagnostic accuracy as the endpoint. Moreover, evaluations need to take into consideration the type of the imaging task (detection and quantification), the properties of the task (lesion size, contrast, edge profile, etc.), and other acquisition and reconstruction parameters.
To evaluate detection tasks, the more acceptable method is observer studies, which involve image preparation, graphical user interface setup, manual detection and scoring, and statistical analyses. Because such evaluation can be time consuming, mathematical models have been proposed to efficiently predict observer performance in terms of a detectability index (d'). However, certain assumptions such as system linearity may need to be made, thus limiting the application of the models to potentially nonlinear IR. For evaluating quantification tasks, conventional method can also be time consuming as it usually involves experiments with anthropomorphic phantoms. A mathematical model similar to d' was therefore proposed for the prediction of volume quantification performance, named the estimability index (e'). However, this prior model was limited in its modeling of the task, modeling of the volume segmentation process, and assumption of system linearity.
To expand prior d' and e' models to the evaluations of IR performance, the first part of this dissertation developed an experimental methodology to characterize image noise and resolution in a manner that was relevant to nonlinear IR. Results showed that this method was efficient and meaningful in characterizing the system performance accounting for the non-linearity of IR at multiple contrast and noise levels. It was also shown that when certain criteria were met, the measurement error could be controlled to be less than 10% to allow challenging measuring conditions with low object contrast and high image noise.
The second part of this dissertation incorporated the noise and resolution characterizations developed in the first part into the d' calculations, and evaluated the performance of IR and conventional filtered backprojection (FBP) for detection tasks. Results showed that compared to FBP, IR required less dose to achieve a threshold performance accuracy level, therefore potentially reducing the required dose. The dose saving potential of IR was not constant, but dependent on the task properties, with subtle tasks (small size and low contrast) enabling more dose saving than conspicuous tasks. Results also showed that at a fixed dose level, IR allowed more subtle tasks to exceed a threshold performance level, demonstrating the overall superior performance of IR for detection tasks.
The third part of this dissertation evaluated IR performance in volume quantification tasks with conventional experimental method. The volume quantification performance of IR was measured using an anthropomorphic chest phantom and compared to FBP in terms of accuracy and precision. Results showed that across a wide range of dose and slice thickness, IR led to accuracy significantly different from that of FBP, highlighting the importance of calibrating or expanding current segmentation software to incorporate the image characteristics of IR. Results also showed that despite IR's great noise reduction in uniform regions, IR in general had quantification precision similar to that of FBP, possibly due to IR's diminished noise reduction at edges (such as nodule boundaries) and IR's loss of resolution at low dose levels.
The last part of this dissertation mathematically predicted IR performance in volume quantification tasks with an e' model that was extended in three respects, including the task modeling, the segmentation software modeling, and the characterizations of noise and resolution properties. Results showed that the extended e' model correlated with experimental precision across a range of image acquisition protocols, nodule sizes, and segmentation software. In addition, compared to experimental assessments of quantification performance, e' was significantly reduced in computational time, such that it can be easily employed in clinical studies to verify quantitative compliance and to optimize clinical protocols for CT volumetry.
The research in this dissertation has two important clinical implications. First, because d' values reflect the percent of detection accuracy and e' values reflect the quantification precision, this work provides a framework for evaluating IR with diagnostic accuracy as the endpoint. Second, because the calculations of d' and e' models are much more efficient compared to conventional observer studies, the clinical protocols with IR can be optimized in a timely fashion, and the compliance of clinical performance can be examined routinely.
Image Quality Evaluation
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 United States License.
Rights for Collection: Duke Dissertations